Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
The value of hydrological information inmultireservoir system operation
Thèse
Jasson Piña Fulano
Doctorat en Génie des EauxPhilosophiæ doctor (Ph. D.)
Québec, Canada
© Jasson Piña Fulano, 2017
The value of hydrological information inmultireservoir system operation
Thèse
Jasson Piña Fulano
Sous la direction de:
Amaury Tilmant, directeur de recherchePascal Côté, codirecteur de recherche
Résumé
La gestion optimale d’un système hydroélectrique composé de plusieurs réservoirs est un problème
multi-étapes complexe de prise de décision impliquant, entre autres, (i) un compromis entre les consé-
quences immédiates et futures d’une décision, (ii) des risques et des incertitudes importantes, et (iii)
de multiple objectifs et contraintes opérationnelles. Elle est souvent formulée comme un problème
d’optimisation, mais il n’existe pas, à ce jour, de technique de référence même si la programmation
dynamique (DP) a été souvent utilisée. La formulation stochastique de DP (SDP) permet la prise en
compte explicite de l’incertitude entourant les apports hydrologiques futurs. Différentes approches
ont été développées pour incorporer des informations hydrologiques et climatiques autres que les ap-
ports. Ces études ont révélé un potentiel d’amélioration des politiques de gestion proposées par les
formulations SDP. Cependant, ces formulations sont applicables aux systèmes de petites tailles en
raison de la célèbre « malédiction de la dimensionnalité ».
La programmation dynamique stochastique duale (SDDP) est une extension de SDP développée dans
les années 90. Elle est l’une des rares solutions algorithmiques utilisées pour déterminer les politiques
de gestion des systèmes hydroélectriques de grande taille. Dans SDDP, l’incertitude hydrologique
est capturée à l’aide d’un modèle autorégressif avec corrélation spatiale des résidus. Ce modèle
analytique permet d’obtenir certains des paramètres nécessaires à l’implémentation de la technique
d’optimisation. En pratique, les apports hydrologiques peuvent être influencés par d’autres variables
observables, telles que l’équivalent de neige en eau et / ou la température de la surface des océans. La
prise en compte de ces variables, appelées variables exogènes, permet de mieux décrire les processus
hydrologiques et donc d’améliorer les politiques de gestion des réservoirs. L’objectif principal de ce
doctorat est d’évaluer la valeur économique des politiques de gestion proposées par SDDP et ce pour
diverses informations hydro-climatiques.
En partant d’un modèle SDDP dans lequel la modélisation hydrologique est limitée aux processus
Makoviens, la première activité de recherche a consisté à augmenter l’ordre du modèle autorégressif et
à adapter la formulation SDDP. La seconde activité fut dédiée à l’incorporation de différentes variables
hydrologiques exogènes dans l’algorithme SDDP. Le système hydroélectrique de Rio Tinto (RT) situé
dans le bassin du fleuve Saguenay-Lac-Saint-Jean fut utilisé comme cas d’étude. Étant donné que ce
système n’est pas capable de produire la totalité de l’énergie demandée par les fonderies pour assurer
pleinement la production d’aluminium, le modèle SDDP a été modifié de manière à considérer les
iii
décisions de gestion des contrats avec Hydro Québec. Le résultat final est un système d’aide à la
décision pour la gestion d’un large portefeuille d’actifs physiques et financiers en utilisant diverses
informations hydro-climatiques. Les résultats globaux révèlent les gains de production d’énergie
auxquels les opérateurs peuvent s’attendre lorsque d’autres variables hydrologiques sont incluses dans
le vecteur des variables d’état de SDDP.
iv
Abstract
The optimal operation of a multireservoir hydroelectric system is a complex, multistage, stochastic
decision-making problem involving, among others, (i) a trade-off between immediate and future con-
sequences of a decision, (ii) considerable risks and uncertainties, and (iii) multiple objectives and
operational constraints. The reservoir operation problem is often formulated as an optimization prob-
lem but not a single optimization approach/algorithm exists. Dynamic programming (DP) has been
the most popular optimization technique applied to solve the optimization problem. The stochastic
formulation of DP (SDP) can be performed by explicitly considering streamflow uncertainty in the
DP recursive equation. Different approaches to incorporate more hydrologic and climatic information
have been developed and have revealed the potential to enhance SDP- derived policies. However, all
these techniques are limited to small-scale systems due to the so-called curse of dimensionality.
Stochastic Dual Dynamic Programming (SDDP), an extension of the traditional SDP developed in the
90ies, is one of the few algorithmic solutions used to determine the operating policies of large-scale
hydropower systems. In SDDP the hydrologic uncertainty is captured through a multi-site periodic au-
toregressive model. This analytical linear model is required to derive some of the parameters needed to
implement the optimization technique. In practice, reservoir inflows can be affected by other observ-
able variables, such snow water equivalent and/or sea surface temperature. These variables, called
exogenous variables, can better describe the hydrologic processes, and therefore enhance reservoir
operating policies. The main objective of this PhD is to assess the economic value of SDDP-derived
operating policies in large-scale water systems using various hydro-climatic information.
The first task focuses on the incorporation of the multi-lag autocorrelation of the hydrologic variables
in the SDDP algorithm. Afterwards, the second task is devoted to the incorporation of different exoge-
nous hydrologic variables. The hydroelectric system of Rio Tinto (RT) located in the Saguenay-Lac-
Saint-Jean River Basin is used as case study. Since, RT’s hydropower system is not able to produce
the entire amount of energy demanded at the smelters to fully assure the aluminum production, a
portfolio of energy contacts with Hydro-Québec is available. Eventually, we end up with a decision
support system for the management of a large portfolio of physical and financial assets using various
hydro-climatic information. The overall results reveal the extent of the gains in energy production that
the operators can expect as more hydrologic variables are included in the state-space vector.
v
Contents
Résumé iii
Contents vi
List of Tables vii
List of Figures viii
Acknowledgments ix
Preface x
Introduction 10.1 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.2 Specific objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1 Literature Review 41.1 The reservoir operation problem . . . . . . . . . . . . . . . . . . . . . . . . . . 41.2 The main solution strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Methods 122.1 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122.2 Optimization problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122.3 Linear Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132.4 Stochastic Dynamic Programming . . . . . . . . . . . . . . . . . . . . . . . . . 142.5 Hydrologic information and large scale systems . . . . . . . . . . . . . . . . . . 152.6 Stochastic Dual Dynamic Programming . . . . . . . . . . . . . . . . . . . . . . 172.7 Climatological information as exogenous variable . . . . . . . . . . . . . . . . . 252.8 Hydropower scheduling and contract management . . . . . . . . . . . . . . . . 252.9 SDDPX formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3 Case study 313.1 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313.2 The Gatineau River Basin and hydropower system . . . . . . . . . . . . . . . . . 313.3 The Saguenay-Lac-Saint Jean River Basin and Rio Tinto system . . . . . . . . . 33
4 Overview of Results 384.1 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
vi
4.2 Incorporation of multi-lag autocorrelation . . . . . . . . . . . . . . . . . . . . . 384.3 Incorporation of exogenous variables . . . . . . . . . . . . . . . . . . . . . . . 434.4 Joint optimization of physical and financial assets . . . . . . . . . . . . . . . . . 45
Conclusion 50
Future work 52
Bibliography 53
A Paper I: Horizontal Approach to asses the Impact of Climate Change on WaterResources Systems 60
B Paper II:Optimizing multireservoir system operating policies using exogenous hy-drologic variables 72
C Autoregressive modeling 88C.1 Multisite periodic autoregressive model MPAR . . . . . . . . . . . . . . . . . . 88
vii
List of Tables
3.1 Gatineau hydropower system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323.2 Rio Tinto hydro-power system characteristics . . . . . . . . . . . . . . . . . . . . . 35
4.1 Extrapolation techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424.2 SDDP formulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444.3 Average annual results - Differences with respect to the SDDP(1) model . . . . . . . 454.4 Average annual results - Differences with respect to the SDDP(1) model . . . . . . . 47
C.1 Autoregressive parameters φp,t and polynomial characteristic roots ui . . . . . . . . . 95
viii
List of Figures
1.1 Illustration of reservoir system optimization as sequential decision process. Modifiedfrom Labadie (2004) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Decision tree in reservoir operation problem (adapted from Pereira et al. (1998)) . . 51.3 Reservoir optimization classification (Ahmad et al., 2014) . . . . . . . . . . . . . . . 51.4 SDP principle when maximizing the sum of immediate and future benefits functions 61.5 Future benefit function - FBF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.6 Dynamic programing Optimization procedure. Adapted from Labadie (2004) . . . . 91.7 SDP and SDDP principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101.8 SDDP and exogenous variables principles . . . . . . . . . . . . . . . . . . . . . . . 11
2.1 Multistage decision problem scheme . . . . . . . . . . . . . . . . . . . . . . . . . . 132.2 Piecewise linear approximation of FBF function Ft+1 . . . . . . . . . . . . . . . . . 172.3 Example reservoir system configuration and connectivity matrix (adapted from Labadie
(2004)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.4 Normalized Convex-Hull approximation at Passes-Dangereuses power station- Rio
Tinto system, Quebec, Canada . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.5 Backwards openings and the approximation of the FBF (adapted from Tilmant and
Kelman (2007)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242.6 The ARX model structure (adapted from Ljung (1999)) . . . . . . . . . . . . . . . . 252.7 Schematization of energy trade . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.8 Schematic SDDPX toolbox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1 Gatineau River Basin and hydropower system . . . . . . . . . . . . . . . . . . . . . 323.2 Weekly average (µ) and standard deviation (σ ) of inflow series - Gatineau River Basin 333.3 Rio Tinto hydropower system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343.4 Weekly average and standard deviation of inflow series - Saguenay-Lac-St-Jean River
Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353.5 Schematic of minimum and maximum weekly levels at Lac-Saint Jean . . . . . . . 363.6 Weekly average and standard deviation of Snow Water Equivalent and precipitation -
Saguenay-Lac-St-Jean River Basin . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.1 Draw-down refill cycle Baskatong and Cabonga Reservoirs. Lag-1 SDDP (Left panel)and multi-lag SDDP(p) (right panel) . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 (a) Statistical distribution of annual spillage losses (b) spillage deviation respect SDDP(1)(c) Statistical distribution of annual energy production (d) energy production deviationwith respect SDDP(1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3 Gatineau and Du Lievre river basins . . . . . . . . . . . . . . . . . . . . . . . . . . 41
ix
4.4 Annual energy generation (a) cumulative distribution functions (b) relative differencesbetween the distribution functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.5 Baskatong drawdown-refill cycle - Climate change and current conditions . . . . . . 434.6 Accumulated drawdown-refill cycle Passes Dangereuses and Lac-Saint-Jean reservoirs 454.7 Statistical distribution of the annual spillage losses (left panel) and the marginal value
of water (right panel) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464.8 Statistical weekly distribution of energy purchases from the portfolio of contracts . . 484.9 % of difference in the power efficiency respect SDDP(1) formulation for both config-
uration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
C.1 Mean and standard deviation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93C.2 (1) Partial correlogram PACF (2) Mean square error MSE (3) AIC (4) BIC . . . . . . 95C.3 Cabonga and Baskatong . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96C.4 Paugan and Chelsea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
x
Acknowledgments
I would like to acknowledge various people who have been part of this adventure.
Firstly, I would like to thank my advisor, Amaury Tilmant for guiding and supporting me over the
years of my Ph.D study. I thank his time, patience and motivation in all the time of research and
writing of this thesis.
I thank also my co-supervisor, Pascal Côté for his kind support, knowledge and generosity.
I would like to thank my thesis committee members, Professors François Anctil, Fabian Bastin and
Marcelo Oliveros, for all of their valuable comments.
Special thanks to my friends in these latitudes: Nicolas and his patience with my French; Charles, his
chocolates and technical discussions; Thibaut and his patience with my Spanish. I would also like to
thank Charles-Hubert for his support and motivation. Thanks to Alex, Hector, Coraline, Bruno, Béné,
Sara, Diane, Maria Natalia, who made this experience more enjoyable.
No podría olvidar mi amada familia que desde la distancia fue un apoyo incondicional. Mil palabras
de agradecimiento por estar junto a mi en todos mis proyectos: gracias Pa, Ma, Adri y Santi. También
a mis amigos que siempre me acompañaron desde Colombia con memes, fotos, videos: Jenny, Moni,
Raquel, Sary, Yiyi, Stiwi, Sergio, Oscar, Maria Cristina...
xi
Preface
This PhD thesis presents the research carried out between January 2014 and August 2017 at the De-
partment of Civil and Water Engineering (Université Laval-UL). This research took place within the
framework of a NSERC-CRD grant with Rio Tinto (RT), under the supervision of Prof. Amaury
Tilmant (UL) and co-supervision of Dr. Pascal Côté (RT). The thesis is based on the following publi-
cations/presentations:
Paper I J. Pina, A. Tilmant, F. Anctil. Horizontal approach to assess the impact of climate change
on water resources systems. Journal of Water Resources, Planning and Management. Published
2016: Pina et al. (2016). http://ascelibrary.org/doi/abs/10.1061/(ASCE)WR.1943-5452.
0000737.
Oral presentation at AGU Fall Meeting 2015: A spatial extrapolation approach to assess the impact
of climate change on water resource systems. https://agu.confex.com/agu/fm15/webprogram/
Paper62891.html
Paper II J. Pina, A. Tilmant, P. Côté. Optimizing multireservoir system operating policies using
exogenous hydrologic variables. Journal of Water Resources Research. Published: Pina et al. (2017).
http://onlinelibrary.wiley.com/doi/10.1002/2017WR021701/abstract
Oral Presentation at EWRI-2017: Optimizing multireservoir system operating policies using exoge-
nous hydrologic variables. https://eventscribe.com/2017/ASCE-EWRI/fsPopup.asp?Mode=
presInfo&PresentationID=254515
Oral Presentation at EGU-2017: Valuing physically and financially-induced flexibility in large-scale
water resources systems. http://meetingorganizer.copernicus.org/EGU2017/EGU2017-9687.
xii
Introduction
Reservoirs are essential for domestic and industrial uses, irrigated agriculture, energy production,
etc. Even though the reservoir operation problem has been studied for decades, its solution remains
challenging. Reservoir operating policies specify the amount of water to be released during a given
stage (time period), or the target storage to be reached at the end of that stage. The complexity of the
problem lies in the interdependence that exists between the immediate and the future consequences
of a release decision. In other words, a balance must be found between outflow, i.e. depleting the
reservoir, and keeping the water in storage for future uses. Since the future inflows are uncertain, the
problem is essentially stochastic. Furthermore, when hydroelectric systems are analyzed, the problem
is nonlinear because the hydropower production function is proportional to the product between the
head (storage) and the releases through the turbines.
The examination of the scientific literature reveals that reservoir operation is often formulated as an op-
timization problem but that no single optimization approach/algorithm exists. Dynamic programming
(DP) and its extensions have been extensively used to solve the reservoir operation problem. The basic
idea behind DP is to decompose the complex problem in a collection of simpler subproblems which
are then solved recursively. The fact that DP can be expanded to account for the hydrologic stochas-
ticity is also an interesting feature when dealing with reservoir operation problems. The stochastic
extension, called Stochastic DP (SDP), performs an optimization on all discrete combinations of the
state variables (storage and hydrologic). Using interpolation techniques, these optimal solutions are
generalized to other points of the state-space domain. In an attempt to better describe the hydrologic
processes, different approaches and extensions of SDP, such as Sampling SDP (SSDP) (Kelman et al.,
1990) and Bayesian SDP (BSDP) (Karamouz and Vasiliadis, 1992) have been developed. However,
those improvements quickly hit the wall: as the traditional DP-based solution strategy relies on the
discretization of the state-space, the problem becomes quickly intractable due to the so-called curse of
dimensionality. Since the computational effort increases exponentially with the number of state vari-
ables, the researchers and practitioners were left with a inevitable trade-off between system complexity
(the number of individual reservoirs) and hydrologic complexity (the number of hydrologic processes
that can be considered). System complexity is desirable to identify synergies between power stations
and to avoid the difficulties associated with aggregation/disaggregation techniques. Hydrologic com-
plexity, on the other hand, should ultimately yield better release policies by reducing the uncertainty
regarding future inflows.
1
This thesis attempts at removing this trade-off using an alternative technique that is not affected by the
curse of dimensionality: Stochastic Dual Dynamic Programming (SDDP). In SDDP the optimal solu-
tion is extrapolated to the rest of the state-space domain from a limited number of discrete points that
are carefully sampled (whereas traditional techniques rely on a dense grid covering the domain and
the solutions are interpolated). SDDP has been used in hydropower-dominated systems such as Nor-
way (Rotting and Gjelsvik, 1992; Mo et al., 2001; Gjelsvik et al., 2010), South and Central America
(Pereira, 1989; Homen-de Mello et al., 2011; Shapiro et al., 2013), New Zealand (Kristiansen, 2004).
The SDDP algorithm also constitutes the core of generic hydro-economic models that have been used
to analyze a variety of policy issues in river basins: e.g. Euphrates-Tigris River basin (Tilmant et al.,
2008), the Nile River basin (Goor et al., 2011), the Zambezi River basin (Tilmant and Kinzelbach,
2012) or in Spain (Pereira-Cardenal et al., 2016; Macian-Sorribes et al., 2016).
In the traditional SDDP formulation, the hydrologic uncertainty is captured through a multi-site pe-
riodic autoregressive model (MPAR). This model is required to analytically derive the extrapolating
functions, and to synthetically generate the different scenarios for the simulation phase of the algo-
rithm. Recent works such us Lohmann et al. (2015), Pritchard (2015), Poorsepahy-Samian et al.
(2016), and Raso et al. (2017) reveal a particular interest in improving the built-in hydrologic model.
This research work follow this trend but focuses on the incorporation of various hydro-climatic infor-
mation into SDDP. To achieve this, additional, exogenous hydrologic state variables must be included
in the state vector, and the built-in hydrologic model must be extended to a multi-site autoregressive
model with exogenous variables (MPARX). This in turn requires that the analytical formulations of
the extrapolating functions be adjusted to accommodate the new exogenous hydrologic variables.
0.1 Objectives
The main objective is to assess the economic value of SDDP-derived operating policies in large-scale
water systems for various hydro-climatic information.
0.2 Specific objectives
1. Develop various analytical formulations for the hydrological model embedded in the SDDP
algorithm, and the respective formulation to determine the approximation of the benefit-to-go
functions.
2. Evaluate the economic performances of the multireservoir system associated with the hydrolog-
ical models developed in (1).
3. Assess the impact of alternative hydrologic information on the management of a hydropower
portfolio including physical (hydropower plants) and financial assets (contracts) .
2
To address the challenges posed by these research objectives, three main activities are developed: 1)
Incorporation of multi-lag autocorrelation 2) Incorporation of exogenous variables and 3) Joint opti-
mization of physical and financial assets.
1. Incorporation of multilag autocorrelation
During this activity, the available Markovian SDDP toolbox (Tilmant et al., 2008), in which
the hydrologic uncertainty is captured by a multi-site lag-1 autoregressive model, is modified
to take into account multiple lags. The built-in hydrologic multisite periodic autoregressive
model of order p (MPAR(p)) is capable of analyzing different hydrologic series, estimating
the parameter of the model, selecting the order p of the periodic model, and generating the
set of hydrologic information required to implement the SDDP algorithm. The mathematical
formulation to couple the higher order MPAR(p) model with the SDDP toolbox is presented
in Paper I. Therein, an assessment of climate change scenarios on a large scale hydropower
system system using the modified toolbox is presented.
2. Incorporation of exogenous variables
In this activity, a hydrologic model capable of processing various hydrologic and climatic infor-
mation is coupled with the optimization model. The built-in model now consists on a MPAR(p)
with lag-b exogenous variables MPARX(p,b). The mathematical formulation to derive the ex-
trapolating functions based on the MPARX model is presented in Paper II.
3. Joint optimization of physical and financial assets
For this activity, a portfolio of financial assets (i.e. sale/purchase energy contracts) is included
in SDDP. A new state variable, accounting for the amount of energy remaining in the contracts,
is included in the state-space vector. The mathematical formulation is presented in Paper II.
Therein, a joint optimization of physical (reservoir and plants) and financial (energy contracts)
assets is presented.
0.3 Outline
The present thesis summarizes the main findings of the two research papers. Chapter 1 presents a
literature review on the reservoir operation problem. Chapter 2 describes the solution strategy to
optimize release policies for large multireservoir systems. Chapter 3 presents different case study
where the various new formulations are applied, and in chapter 4 the main findings are summarized.
In chapter 5 can be found the conclusion remarks and in the appendix the research papers.
3
Chapter 1
Literature Review
1.1 The reservoir operation problem
The operation of a multi-reservoir system is a complex, multistage, stochastic decision-making prob-
lem involving, among others, (i) a trade-off between immediate and future consequences of a release
decision, (ii) considerable risks and uncertainties and (iii) multiple objectives and operational con-
straints (Oliveira and Loucks, 1997). The complexity of the problem lies in the interdependence that
exists between the immediate and the future consequences of a decision at a given stage. In other
words, a balance between storage and release decisions must be found at each stage (Figure 1.1).
Stage
t
Stage
T
Inflow1
Stage
1
Release1
Storage1
Benefits1
Storage2
State transition
Inflowt Releaset
Storaget
Benefitst
Storaget+1
InflowT ReleaseT
StorageT
BenefitsT
StorageT+1
Allocation decision
State Returns
Figure 1.1: Illustration of reservoir system optimization as sequential decision process. Modified fromLabadie (2004)
Since the future inflows are uncertain, the problem is essentially stochastic. At each stage of the
decision process, reservoir operators face the future hydrologic uncertainty and the decision made can
affect the availability of the resource and thus the future benefits (Figure 1.2).
This decision making problem has been studied for several decades and a state-of-art reviews can be
4
Operating consequences
Future hydro conditions
Decision
Reservoir
Use reservoir
Dry Deficit
Wet ok
Do not use reservoir
Dry ok
WetSpillage losses
Figure 1.2: Decision tree in reservoir operation problem (adapted from Pereira et al. (1998))
found in Yeh (1985), Labadie (2004), Rani and Moreira (2009) and more recently in Ahmad et al.
(2014). The examination of the scientific literature reveals that reservoir operation is often formulated
as an optimization problem but that no single optimization approach/algorithm exists.
Reservoir Operation
Linear Programming
(LP)
Network Flow programming
Interior Point Method
Non-Linear Programming
(NLP)
Sequential Linear Programming
Sequential Quadratic
Programming
Method of Multiplier
Generalized Reduced
Gradient Method
Dynamic Program (DP)
Deterministic DP
Stochastic DP
Computational Intelligence (CI)
Fuzzy Set Theory
Artificial Neural Network
Evolutionary computer
Figure 1.3: Reservoir optimization classification (Ahmad et al., 2014)
In the stochastic case, however, two techniques exist: Implicit Stochastic Optimization (ISO) and
Explicit Stochastic Optimization (ESO). ISO methods are actually deterministic methods which use
a large number of historical or synthetically generated hydrological scenarios to derive optimal op-
5
eration policies: they use optimization, regression and simulation techniques to refine promising op-
eration rules iteratively. ESO formulations, on the other hand, require the explicit representation of
probabilistic streamflows or other uncertain parameters such us energy and water demands, spot prices
for energy, etc. (Labadie, 2004). To handle the complexity of the reservoir operation problem, a tem-
poral decomposition approach is required (Zahraie and Karamouz, 2004). This hierarchical approach
relies on a chain of optimization models for long, mid and short term planning horizons. The solution
strategies presented herein focus on the mid-term hydropower scheduling, which seeks to determine
optimal weekly release policies. As we will see in the next section, even though various optimization
techniques are available to solve the mid-term reservoir operation problem (Figure 1.3), this research
relies on extension of Dynamic Programming.
Optimal decision
Future benefits function
Immediate benefitsfunction
End-of-period storage
Immediate + future benefits functions
Be
nef
its
$
Marginal water values
Figure 1.4: SDP principle when maximizing the sum of immediate and future benefits functions
1.2 The main solution strategies
Along with linear programming and nonlinear programming, dynamic programming and its exten-
sions have been extensively studied. Dynamic programming (DP), first introduced by Bellman (1957),
solves the problem by breaking the multistage problem into simpler one-stage subproblems, which are
6
then solved recursively. With this principle in mind, the objective function of the one-stage optimiza-
tion problem becomes the sum of the immediate and future benefits from system operation (Figure
1.4). As we can see, the immediate benefits decrease as the end-of-period storage increases as less
water is available for immediate uses. At the same time, future benefits increase as more water is kept
in storage. The derivatives of the immediate and future benefit functions correspond to the immediate
and future marginal water values respectively (Tilmant et al., 2008) and, at the optimal solution, both
values are identical. The marginal value of water indicate what the operator would be willing to pay
to get an additional unit of water in a particular reservoir and at a given time of the year.
To account for the hydrologic uncertainty, DP can be expanded by adding hydrological variables in the
state vector. Consequently, the release decisions are now function of the storage and the hydrologic
state variable (e.g. previous or current inflow). This stochastic DP (SDP) formulation, often referred
to as a Markov decision process, explicitly considers the streamflow lag-1 correlation found in the
flow records; the recursive equation uses the fact that inflow during any given time period is related to
the previous one by a conditional probability (Yeh, 1985).
End-of-period storage
Ben
efit
s $
Stochastic variable
Figure 1.5: Future benefit function - FBF
SDP solves the problem by replacing the continuous domain by a grid, and by solving the one-stage
DP optimization problem at each grid point. These optimal solutions are then generalized the rest of
the domain using interpolation techniques (Figure 1.5). Since this optimization is performed condi-
tionally on all discrete combinations of the state variables, this discrete approach is limited to small
scale problems with no more than four state variables. To illustrate the so-called curse of dimen-
sionality of a four-dimensional problem, let us imagine that the state variables are discretized in 10
values. Then, the one-stage optimization must be evaluated over a grid of 10 × 10 × 10 × 10 points
7
(104). Hence, the computational effort increases exponentially with the number of state variables as
10#variables.
Various strategies for dealing with the dimensionality issue associated with DP have been proposed
in the literature. For example, Turgeon and Charbonneau (1998), Saad et al. (1996), Archibald et al.
(1997) and Archibald et al. (2006) use aggregation- disaggregation techniques to reduce the scale of
the problem. Bellman and Dreyfus (1962) suggest the Dynamic Programming Successive Approxima-
tion (DPSA), which decomposes the multidimensional problem into a sequence of one-dimensional
problem by optimizing over one state variable at a time. Then releases are explicitly obtained from the
mass balance equation as a function of specified beginning and ending storage. Incremental Dynamic
Programming (IDP) and Discrete Differential Dynamic Programming (DDDP) address the dimension-
ality problem by restricting the state space to a corridor around a current given solution (Karamouz
et al., 2003). The methods are highly sensitive to initial storage trajectories and the discretization
intervals must be carefully selected to provide accurate solutions at a reasonable computational time
(Labadie, 2004). Although, these efforts reduce computation time, the curse of dimensionality is not
removed.
Along with the curse of dimensionality, the representation of streamflow persistence and hydrologic
forecasting information is an important issue when applying SDP: better streamflow foresight is ex-
pected to improve reservoir operation because it allows time for better decision making (Georgakakos,
1989). The generation of scenarios and forecast information for water resources management appli-
cations relies on the use of different stochastic hydrologic models (Pagano et al., 2004; Gelati et al.,
2010). Various approaches are available to forecast reservoir inflows, from regression relationships be-
tween inflows and climate observed data (e.g. snowpack, soil moisture, fall and winter precipitation),
to models properly initialized with climate forecast (e.g. Ensemble Streamflow Prediction (ESP) fore-
casts, forecast of ENSO, downscaled numerical climate model forecasts, etc.)(Anghileri et al., 2016;
Georgakakos, 1989). The National Weather Service’s (NWS) ESP procedure (Day, 1985) produces
streamflow forecast in the form of multiple hydrographs, and the forecast of ENSO are currently
available up to a year or more in advance (Gelati et al., 2014).
To exploit the potential value of these forecasts, different extensions of SDP have been developed. For
example, Kelman et al. (1990) proposed a sampling SDP (SSDP) that captures the complex temporal
and spatial structure of the streamflow process through the use of a large number of sample stream-
flows scenarios, instead of assuming that inflow stochasticity in SDP follows a probability density
function. Karamouz and Vasiliadis (1992) developed the Bayesian SDP (BSDP) which uses Bayesian
decision theory to incorporate new information by updating the transition probabilities; in BSDP nat-
ural and forecast uncertainties are both included in the model.
The advantage of using different hydrologic variables in SDP formulations has been presented in sev-
eral references. Bras et al. (1983) presented the introduction of real time forecast with an adaptive
control technique where flow transition probabilities and system objective are continuously updated.
8
Infl
ow
s
Time
DP
Statistical Frequency
analysis
SDP
Probability distributions
FBF
SSDP-BSDP
Statistical Frequency
analysis
Probability distributions
Historical series or
stochastic stream
generation model
Simulation model
Hydrologic Forecasting information
Hydrologic persistenceMonte
Carlo
ESOISO
Fore
cast
in
form
ati
on
Time
`
Figure 1.6: Dynamic programing Optimization procedure. Adapted from Labadie (2004)
Stedinger et al. (1984) developed a SDP model which employs the best inflow forecast of the cur-
rent period to define the policy. Georgakakos (1989) discussed the value of streamflow forecasts in
reservoir operation. Kim and Palmer (1997) compared the performance of the BSDP formulation
and three alternative SDP models, when the seasonal flow forecast and other hydrologic information
are included in the state vector. Faber and Stedinger (2001) and Kim et al. (2007) employed ESP
forecasts and snowmelt volume forecasts using SSDP formulation. Côté et al. (2011), introduced in
SSDP a new hydrological state variable given as a linear combination of snow water equivalent and
soil moisture. More recently, Desreumaux et al. (2014) presented the effect of using various hydro-
logical variables on SDP-derived policies of the Kemano hydropower system in British Columbia.
Anghileri et al. (2016) presented a forecast-based adaptive management framework for water supply
reservoirs and evaluate the contribution of long-term inflow forecasts to reservoir operations.
However, most of the studies described above are limited to small-scale problems, meaning that a
trade-off must be found between the complexity of system to be studied and the complexity of the hy-
drologic processes that can be captured. When analyzing a large-scale system, operators must decide
whether to use a simplification of the system (e.g. aggregating storage capacity), or use fewer hy-
9
drologic state variables to describe the hydrologic process, leading somehow to a loss of information,
and to a loss of spatiotemporal synergies that can be captured when analyzing the whole system. This
trade-off can largely be removed by using stochastic dual DP (SDDP).
SDDP, first introduced by Pereira and Pinto (1991), is one of the few available algorithms to opti-
mize the operating policies of large-scale hydropower systems. The solution approach is based on
the approximation of the expected future benefit function (FBF) of SDP by piecewise linear functions
(Figure 1.7). With SDDP, there is no need to evaluate the FBF over a dense grid as the function
can now be derived from extrapolation (and not interpolation). The accuracy of the approximation is
increased by adding new linear segments through a two-phase iterative algorithm. The set of linear
segments can be interpreted as Benders cuts in a stochastic multistage decomposition algorithm, and
its determination relies on the primal and dual information of the optimal solution of each subproblem.
To implement the efficient decomposition scheme, each nonlinear SDP subproblem must be formu-
lated as a convex problem, such as a linear program (LP). This constitutes the main drawback of the
technique since all the relations associated with the problem, objective function, and constraints, must
be linear.
End-of-period storage
Ben
efit
s $
Stochastic variable
End-of-period storage
Ben
efit
s $
Stochastic variable
Sampling Point 2
Sampling Point 1
Piecewise linear approximation
True function
Hyperplane 1
Hyperplane 2
Figure 1.7: SDP and SDDP principles
SDDP has largely been used in hydropower systems, such as Norway (Rotting and Gjelsvik, 1992;
Mo et al., 2001; Gjelsvik et al., 2010), South and Central America (Pereira, 1989; Homen-de Mello
et al., 2011; Shapiro et al., 2013), New Zealand (Kristiansen, 2004) and Turkey (Tilmant and Kelman,
2007). Some improvements to deal with the nonlinear water head effects have been developed by
using a convex hull approximation of the hydropower function and can be found in Goor et al. (2011)
and Cerisola et al. (2012). The SDDP algorithm constitutes the core of generic hydro-economic
models that have been used to analyze a variety of policy issues in the Euphrates-Tigris River basin
(Tilmant et al., 2008), the Nile River basin (Goor et al., 2011), the Zambezi River basin (Tilmant and
Kinzelbach, 2012) or in Spain (Pereira-Cardenal et al., 2016; Macian-Sorribes et al., 2016).
10
Infl
ow
s
Time
Hydrologic uncertainty
MPAR(p)
SDDP
FBF approximations
Hyd
ro-c
lima
tic
info
rma
tio
n
Time
Hydrologic uncertaintyMPARX(p,b)
SDDPX
Figure 1.8: SDDP and exogenous variables principles
In SDDP, the hydrologic uncertainty is captured through a multi-site periodic autoregressive model.
This model is capable of representing seasonality, serial and spatial streamflow correlations within
a river basin and among different basins. Furthermore, it is required to analytically derived the ap-
proximation of the FBF, and to synthetically generate the different scenarios for the simulation phase.
Recent works reveal a particular interest in improving the built-in hydrologic model. For example,
Lohmann et al. (2015) presented a new approach to include spatial information. Pritchard (2015)
modeled inflows as a continuous process with a discrete random innovation, Poorsepahy-Samian et al.
(2016) proposed a methodology to estimate the cuts parameters when a Box-Cox transformation is
used to normalize inflows, and more recently, Raso et al. (2017) present a streamflow model with a
multiplicative stochastic component and a non-uniform time step.
This PhD thesis fits into this trend and seeks to assess the value of different hydro-climatic information
when operating large-scale water resource systems. To achieve this, additional exogenous hydrologic
state variables are incorporated into the SDDP algorithm. Since the modeling of the hydrologic un-
certainty in SDDP is restricted to linear additive models (Infanger and Morton, 1996; De Queiroz
and Morton, 2013), the natural extension to include climatic variability into the autoregressive model
is the MPAR model with exogenous variables (MPARX) (Figure 1.8). The reader should refer to
Ljung (1999), Ltkepohl (2007) and Hannan and Deistler (2012) for detailed presentations of MPARX
models.
11
Chapter 2
Methods
2.1 Outline
This chapter explains how the exogenous hydrologic variables are incorporated into the SDDP al-
gorithm. It starts with overview of the optimization techniques available for solving the reservoir
operation problem with a particular attention given to Stochastic Dynamic Programming (SDP). Al-
ternative SDP formulations, each employing different hydrologic information, are presented, their
strengths and weaknesses discussed. Finally, the Stochastic Dual Dynamic Programming (SDDP)
algorithm is described.
2.2 Optimization problem
Recall that the reservoir operation problem is a multistage decision-making problem (Figure 2.1).
When it is formulated as an optimization problem, the goal is to determine a sequence of optimal
allocation decisions xt (e.g. reservoir release and spillage, water withdrawals, etc) that maximizes
the expected sum of benefits from system operation Z, over a planning period time T , while meeting
operational and/or institutional constraints. The mathematical formulation of the multistage decision-
making problem can be written as:
Z = max
E
[T
∑t=1
αtbt(St ,qt ,xt)+αT+1ν(ST+1)
] (2.1)
subject to:
12
St+1 = ft(St ,qt ,xt) (2.2)
gt+1(St+1)≤ 0 ∀t (2.3)
at+1(xt+1)≤ 0 ∀t (2.4)
Stage
t
Stage
T
Stage
1 State transition
Allocation decision
State Returns
𝑞𝑡 𝑥𝑡 𝑞𝑇 𝑥𝑇𝑞1 𝑥1
𝑠𝑡 𝑠𝑡+1 𝑠𝑇 𝑠𝑇+1𝑠1
𝑏1(𝑆1, 𝑞1, 𝑟1)
𝑠2
𝑏𝑡(𝑆𝑡, 𝑞𝑡, 𝑟𝑡) 𝑏𝑇(𝑆𝑇 , 𝑞𝑇 , 𝑟𝑇)
Figure 2.1: Multistage decision problem scheme
where bt(·) is the immediate benefit function, ν(·) is the terminal value function, αt is the discount
factor at stage t and E[·] is the expectation operator. In many reservoir operation problems the vector
of the state variables St includes the beginning-of-period storage st and any hydrological variable
ht . For a hydropower-dominated system of J reservoirs and D demand sites for off-stream uses (e.g.
municipal and industrial uses, irrigated agriculture), the immediate benefit function includes the net
benefits from hydropower HPt [$], the benefits from off-stream uses NBt [$], and penalties for not
meeting target water demands and/or violating operating constraints:
bt(·) = HPt +NBt −ξᵀt zt (2.5)
where zt (J× 1) is the vector of slack variables with the violations of operational constraints (e.g.
energy deficit, environmental flows, etc.) which are penalized in the objective function by the vector
ξt (J×1) of penalties [$/unit].
2.3 Linear Programming
Linear Programming (LP) is one of the most popular optimization techniques applied in water re-
sources management. Its attractiveness lies in ability to handle large scale problems, to converge to a
global solution, to allow for a sensitivity analysis from the duality characteristics of linear program-
ming, and the availability of generic softwares for solving LP problems. The main disadvantage is
13
the fact that the LP requires that all the relations associated with the problem, objective function and
constraints, to be linear or linearizable. Moreover, if the stochastic formulation of LP is rarely im-
plemented because of the computational burden, the two-stage stochastic linear programming with
recourse, first introduced independently by Dantzig (1955) and Beale (1955), can be used for solving
the problem deterministically for each of the the several scenarios of future inflows. This formulation
leads to an extremely large-scale linear programming problem, which can be reduced by utilizing Ben-
ders decomposition. The basic idea is to express the expected value of the second stage by a scalar and
to replace the second-stage constraints sequentially by cuts, which are necessary conditions expressed
only in terms of the first stage variables (Infanger, 1993). Another decomposition strategy for solving
large-scale stochastic programs is progressive hedging (PH) (Rockafellar and Wets, 1991) which is a
scenario-based decomposition technique. PH has been used as an effective heuristic technique for ob-
taining approximate solutions to multistage stochastic programs (Hart et al., 2012; Carpentier et al.,
2013)
2.4 Stochastic Dynamic Programming
Next to linear programming, Dynamic programming (DP) has been the most popular optimization
technique applied to water resources planning and management. The method was first introduced by
Bellman (1957) and solves the problem (Equations 2.1 to 2.4) by breaking the multistage problem
into simpler subproblems over each stage, which are solved recursively. DP can handle non-linear
relationships and discontinuous functions. DP performs an optimization on all discrete combinations
of the state variables. These optimal solutions are generalized for other points of the state variables
by a continuous function, using an interpolation approach (e.g. linear, cubic spline) (Johnson et al.,
1993; Tejada-Guibert et al., 1993; Kitanidis et al., 1987). Stochastic DP (SDP), often referred to a
Markov decision process, solves the problem by discretizing stochastic variables, as well as the system
status, to obtain an optimal policy for each discrete value of the reservoir system (Rani and Moreira,
2009). If the vector of state variables St includes the beginning-of-period storage st and any choice
of hydrologic state variable ht , the recursive SDP equation can be written as (Tejada-Guibert et al.,
1995):
Ft(st ,ht) = Eqt |ht
[maxxt{αtbt(·)+ E
ht+1|ht ,qt
[αt+1Ft+1(st+1,ht+1)]}] (2.6)
where E[·] is the expectation operator to observe hydrological condition ht+1 given the hydrological
state ht , and it is obtained from the conditional probabilities P(ht+1 | ht). SDP directly incorporates
both the probability distributions of random variables and the temporal persistence between successive
flows through the use of flow transition probabilities. Since the expectation operator acts on the max-
imization, the release decisions are made after the hydrologic variable ht is known (Tejada-Guibert
et al., 1995).
14
When the temporal persistence is not modeled, the hydrologic variable is not included in the SDP and
the equation 2.6 is reduced to the deterministic DP equation 2.7. The Pseudo-code description of the
DP procedure is presented in Algorithm 1.
Ft(st) = maxxt{αtbt(·)+αt+1Ft+1(st+1)} (2.7)
Initialize the FBF of the last stage;FT (sT+1) = 0;
for t = T,T −1, ...,1 dofor each storage level st = s1
t , ...,sKnt do
solve the one-stage problem Equation 2.7Ft(st) = max
xt{αtbt(·)+αt+1Ft+1(st+1)}
s.t operational constraints equations 2.2 to 2.4endCreate a complete FBF, Ft(st), for the previous stage by interpolating the values{
Ft(sknt ),kn = 1, ...,Kn
}end
Algorithme 1 : Pseudo-code description of the DP procedure
The most common choices for hydrological state variable have been the current flow qt and the pre-
vious flow qt−1. When the current flow is used the expectation Eqt |ht can be omitted and equation 2.6
becomes:
Ft(st ,qt) = maxxt{αtbt(·)+ E
qt+1|qt
[αt+1Ft+1(st+1,qt+1)]} (2.8)
Likewise, if the state hydrological variable is the previous flow equation 2.6 becomes:
Ft(st ,qt−1) = Eqt |qt−1
[maxxt{αtbt(·)+αt+1Ft+1(st+1,qt)}] (2.9)
and the Pseudo-code description of the SDP procedure is presented in algorithm 2.
2.5 Hydrologic information and large scale systems
Among several concerns, the representation of streamflow persistence and hydrologic forecasting in-
formation in the decision process is a critical issue when applying SDP (Labadie, 2004; Kelman et al.,
1990). Thus, better hydrologic information, incorporated as state variables, has the potential for en-
hancing SDP-derived policies. For example, Bras et al. (1983) presented the introduction of real time
forecast with an adaptive control technique where flow transition probabilities and system objective
are continuously updated over finite transient periods before achieving steady state conditions. Ste-
dinger et al. (1984) developed a SDP model which employs the best inflow forecast of the current
15
Initialize the FBF of the last stage;FT (sT+1,qT ) = 0;
for t = T,T −1, ...,1 dofor each storage level st = s1
t , ...,skmt , ...,sKm
t dofor each past inflow qt−1 = q1
t−1, ...,qkkt−1, ..., ,q
Kkt−1 do
for each inflow scenario of stage t conditioned to the past inflow scenarioqt = q1
t , ...,qk jt , ...,qK j
t dosolve the one-stage problem Equation 2.9 considering an initial storage skm
t and theinflow scenario qk j
t
Fk jt (skm
t ,qkkt−1) = max
xt{αtbt(·)+αt+1Ft+1(st+1,q
k jt )}
s.t operational constraints equations 2.2 to 2.4endcalculate the expected value of the benefits obtained across the conditioned inflowscenarios
Ft(skmt ,qkk
t−1) = ∑k j
(P
k j|kk
Fk jt (skm
t ,qkkt−1)
)endCreate a complete FBF, Ft(st ,qt−1), for the previous stage by interpolating the values{
Ft(skmt ,qkk
t−1),km = 1, ...,Km;kk = 1, ...,Kk
}end
endAlgorithme 2 : Pseudo-code description of the SDP procedure
period to define the policy. Kelman et al. (1990) proposed a sampling SDP (SSDP) to better capture
the complex temporal and spatial structure of the streamflow process. SSDP uses a large number
of streamflows scenarios, instead of assuming that inflow stochasticity in SDP follows a probability
density function. SSDP implementation relies on two different models: i) a decision model which
chooses an optimal release that maximizes the future and current benefit for each stage, state, and
scenario taking into account the streamflow uncertainty, and ii) a simulation model which uses the
optimal releases for each scenario to update the future value maintaining a realistic description of
streamflow series.
Karamouz and Vasiliadis (1992) developed the Bayesian SDP (BSDP) which includes inflow, storage,
and forecast as state variables. BSDP describes streamflows with a discrete lag-1 Markov process, and
uses Bayesian decision theory to incorporate new information by updating the transition probabilities.
In BSDP, natural and forecast uncertainties are both captured in the model. When the current inflow
qt and the seasonal or monthly flow forecast ft+1 are used as hydrologic state variables equation 2.6
becomes:
Ft(st ,qt , ft+1) = Eqt |qt , ft+1
[maxxt{αtbt(·)+ E
qt+1 ft+2|qt , ft+1,qt
[αt+1Ft+1(st+1,qt+1, ft+2)]}] (2.10)
Depending on the selection of the current hydrologic state variables ht and the next state hydrological
16
variables ht+1, the conditional expectations in Equation 2.10 can be determined for four different cases
(further details see Kim and Palmer (1997)). Therein, the potential advantage of using the seasonal
flow forecast and other hydrologic information is illustrated by comparing the performance of the
BSDP formulation and three alternative SDP models.
As mentioned earlier, since the optimization is performed conditionally on all discrete combinations
of the state variables, the problem gives rise to the curse of dimensionality: assuming k state variables,
discretized into N values the computational effort required to solve equation 2.6 increases exponen-
tially with the number of reservoirs J as(
NkJ)
. Then, if a large-scale system is studied using SDP,
a balance must be found between the complexity of the system (e.g number of reservoirs) and the
complexity of the hydrologic process that can be captured. Although strategies such as aggregating
the storage capacity or using fewer hydrologic variables can be implemented, they can somehow lead
to a loss of information, and to a loss of spatiotemporal synergies that can be captured when analyzing
the whole system. This trade-off between system and hydrologic complexities can largely be removed
by using stochastic dual DP (SDDP).
2.6 Stochastic Dual Dynamic Programming
Stochastic dual DP (SDDP) was first proposed by Pereira and Pinto in 1991. SDDP is not affected by
the curse of dimensionality and can therefore be used to optimize the operating policies of large-scale
hydropower systems. The solution approach is based on the approximation of the FBF functions of
SDP by piecewise linear functions (Figure 2.2). There is no need to evaluate Ft+1 over a dense grid as
the function can now be derived from extrapolation (and not interpolation).
Sampling Point 2
Sampling Point 1
Piecewise linear approximation
True function
Hyperplane 1
Hyperplane 2
𝑠𝑡+1
𝑞𝑡
𝐹𝑡+1
Figure 2.2: Piecewise linear approximation of FBF function Ft+1
17
With SDDP the multistage optimization problem (Equations 2.1 to 2.4) can be broken into a series of
one-stage linear programming (LP) problems which are solved recursively. At stage t, describing the
state of the system with the storage st and using as hydrologic state variable the previous inflow qt−1,
the recursive equation can be written as:
Ft(st ,qt−1) = max{αtbt(·)+αt+1Ft+1} (2.11)
The problem is bounded by L Bender’s cuts which are represented by the inequality constraints:
Ft+1−ϕ
lt+1
ᵀst+1 ≤ γ
lt+1
ᵀqt +β
lt+1
...
Ft+1−ϕLt+1
ᵀst+1 ≤ γ
Lt+1
ᵀqt +β
Lt+1
(2.12)
ϕt+1 (J×1), γt+1 (J×1), βt+1 are the linear parameters of the approximated FBF Ft+1. Likewise, the
stage to stage transformation function (i.e. the mass balance equation) is expressed as:
st+1−CMR(rt + lt)+ et(st ,st+1) = st +qt (2.13)
where st+1 is the vector (J× 1) of storage at the end of the period, rt is the vector (J× 1) of the
turbined flows, lt and et are the vectors (J× 1) of spillage and evaporation losses respectively, CMR
is the reservoir system connectivity matrix, CMRj,k= 1(-1) when reservoir j receives (releases) water
from (to) reservoir k. Figure 2.3 displays an example of a reservoir system configuration and the
connectivity matrix.
The linear segments Ft+1 are obtained from the dual solutions of the optimization problem at each
stage and can be interpreted as Benders cuts in a stochastic, multistage decomposition algorithm.
SDDP uses an iterative optimization/simulation strategy to increase the accuracy of the solution by
adding new cuts. To implement the decomposition scheme, the one-stage optimization problem must
be formulated as a convex problem, such as a linear program. Nonetheless, the power generation
function depends on the product of the turbined outflow and the net head on the turbine:
Pt = η(st ,st+1,rt) ·ρ ·g · rt ·ht(st ,st+1,rt) (2.14)
where Pt[W] is the power produced in the plant, η is the overall efficiency of the power plant, ρ
[kg/m3] is the density of water, g [m/s2] is the acceleration due to gravity, rt [m3/s] is the release
through the turbines and ht [m] is the net head which is a non linear function depending on the storage
levels at the begening, st and at the end st+1 of the period and the head losses.
To deal with the head effects on the hydropower production function Pt , a convex hull approximation
is stored in the constraints set (2.15). The linear parameters ψ , ω and δ are determined using the
procedure described in Goor et al. (2011).
18
1
2
3
4
Reservoir
Power Plant
𝐶𝑀𝑅 =
−1 0 0 00 −1 0 00 +1 −1 0+1 0 +1 −1
𝑞1
𝑞2
𝑞3
𝑞4
𝑟1
𝑟2
𝑟3
𝑟4
Figure 2.3: Example reservoir system configuration and connectivity matrix (adapted from Labadie(2004))
Pt −ψ
1st+1/2−ω1rt ≤ ψ
1st/2+δ1
...
Pt −ψHst+1/2−ω
Hrt ≤ ψHst/2+δ
H
(2.15)
Then, the immediate benefits of equation 2.11 are calculated as:
bt(st ,qt ,st+1,rt) =J
∑j=1
(πh( j)−θh( j))Pt( j)τt −ξ
ᵀt zt (2.16)
where τt is the number of hours in period t, π is the energy price [$/Wh] and θ is the operation and
maintenance cost [$/Wh]. As defined in Equation 2.5, zt is the vector of slack variables, penalized in
the objective function by the vector ξt of penalties [$/unit].
The decision variables such as storage st+1, releases rt and spillage losses lt are limited by lower and
upper boundaries:
st+1 ≤ st+1 ≤ st+1
rt ≤ rt ≤ rt
lt ≤ lt ≤ lt (2.17)
19
Figure 2.4: Normalized Convex-Hull approximation at Passes-Dangereuses power station- Rio Tintosystem, Quebec, Canada
2.6.1 Future Benefit Function Approximation
In SDDP, the hydrologic uncertainty is typically captured through a multi-site periodic autoregres-
sive model (MPAR). This model is capable of representing seasonality, serial and spatial stream-flow
correlations within a river basin and among different basins. It is also needed to analytically derive
the FBF approximations, and to produce synthetic streamflows scenarios for the simulation phase of
the iterative procedure. Furthermore, the convexity requirement of SDDP is guaranteed because the
MPAR is linear (Infanger and Morton, 1996; De Queiroz and Morton, 2013).
Autoregressive Model
Autoregressive (AR) models have been extensively used in hydrology and water resources since
1960’s. Its popularity and attractiveness rely on the simplest formulation and its intuitive type of time
dependence: variables at time t are dependent on the preceding ones. First introduced by Thomas
and Fiering (1962) and later by Box and Jenkins (1970), AR models can be represented as models
with constant parameters, parameters varying with time and combination of both. The models with
constant parameters are often implemented for modeling annual time series. Models with periodic
parameters are usually used with time series of intervals that are fraction of the year (e.g. seasons,
months, weeks, etc.) Salas et al (1980). The latter models are referred to as periodic AR (PAR)
20
models, and the periodicity may be in the mean, variance and/or autoregressive parameters. Since the
reservoir operation analysis involves time series at various geographic locations, spatial correlation
is also required. Then, at each site j the hydrologic process can be modeled by a multi-site periodic
autoregressive MPAR model of order p represented by:
(qt( j)−µqt ( j)
σqt ( j)
)=
p
∑i=1
φi,t( j)
(qt−i( j)−µqt−i( j)
σqt−i( j)
)+ εt( j) (2.18)
where qt is the time dependent variable for year v and time t, with t=1,2,...,52 weeks. µqt and σqt are
the periodic mean and standard deviation of qt , respectively, φi,t( j) are the autoregressive parameters
of the p order periodic model, and εt( j) is a time independent-spatially correlated stochastic noise.
Assuming that the noise εt( j) follows a 3-parameters (µv( j), σv( j) and κt( j)) log normal distribution:
fεt( j) =1
(εt( j)−κt( j))√
2πσv( j)e−0.5
(log(εt ( j)−κt ( j))−µv( j)
σv( j)
)2
(2.19)
with mean µεt , variance σ2ε,t
µεt ( j) = κt( j)+ e
(µv( j)+ σ2
v ( j)2
)
σ2ε,t( j) = e2(µv( j)+σ2
v ( j))+ e(2µv( j)+σ2v ( j)) (2.20)
the lower bound κt( j) which ensure non-negative inflows qt( j) > 0, is defined from equation 2.18 is
defined as:
εt( j)>−µqt ( j)σqt ( j)
−p
∑i=1
φi,t( j)
(qt−i( j)−µqt−i( j)
σqt−i( j)
)= κt( j) (2.21)
and the parameters µv( j) and σv( j) determined as:
µv( j) = 0.5Log σ2ε,t
Λ( j)(Λ( j)−1) (2.22)
σv( j) =√
log(Λ( j)) (2.23)
Λt( j) = 1+ σ2ε,t
κ2t ( j) (2.24)
the standardized stochastic noise Vt is estimated as:
Vt( j) =log(εt( j)−κt( j))−µv( j)
σv( j)(2.25)
21
The spatial statistical dependence of reservoir inflows is introduced by the lag0-covariances and cross
variances of the standardized stochastic noise. Then for the collection of independent standardized
noises Vt of each node j of the system Vt = [Vt(1), . . .Vt(J)], the spatial model can be written as:
Vt = AtWt (2.26)
where Wt is a column vector of J independent elements consisting of white noises, normally distributed
with zero mean and variance equal to 1. The estimation of matrix A can be obtained from the Cholesky
factorization of the covariance matrix of standardized noise at each node j:
At Aᵀt =Cov(Vt) (2.27)
SDDP and MPAR(1)
To derive the mathematical formulation of the FBF approximation let us assume an autoregressive
model of order p = 1, then equation 2.18 can be written as:
qt( j)−µqt ( j)σqt ( j)
= φt( j)
(qt−1( j)−µqt−1( j)
σqt−1( j)
)+ εt( j) (2.28)
SDDP uses a two phases strategy to increase the accuracy of the solution by adding new cuts: a
backward optimization and a forward simulation. Both phases require different sets of inflows. In the
backward phase, K inflows scenarios (backward openings) at each node of the system are generated
by the using the MPAR model. These scenarios are needed to analytically calculate the hyperplanes’
parameters, and ultimately to derive the upper bound to the true expected FBF. In the forward phase,
the MPAR model generates M synthetic reservoir inflows sequences to simulate the system behavior
over the planning period.
The calculation of the linear parameters ϕ lt , γ l
t and β lt of the approximated Ft+1 (equation 2.12) relies
on the primal and dual information available at the optimal solution. Let us say that at stage t, s◦t and
q◦t−1 are sampled and, in order to include the stochasticity of the problem, the K vectors of inflows
qKt are generated. The one-stage SDDP subproblem 2.11 to 2.17 is solved for K reservoir inflow
branches qkt . The expected FBF Ft+1, stored in the form of cuts, is the expected value of the K FBF
Fkt+1 calculated for each inflow branch (Figure 2.5) (Tilmant and Kelman, 2007).
According to the Kuhn-Tucker conditions for optimality the derivative of the objective function with
respect to the state variables S is given by:
∂F∂Si
= ∑λi∂gi
∂Si(2.29)
22
being λi the dual information of the optimization problem and gi the linear constraints 2.13 to 2.15.
Then, the slopes ϕl,kt and γ
l,kt of the functions Fk
t
Fkt ≤∑
Jϕ
l,kt ( j)s◦t ( j)+∑
Jγ
l,kt ( j)q◦t−1( j)+β
l,kt (2.30)
which will be added to the expected cost-to-go function at stage (t−1), can be calculated as:
∂Fkt
∂ st( j) = ϕ
l,kt ( j) = λ
kw,t( j)+
H
∑h=1
λh,khp,t( j)ψh
t+1( j)/2 (2.31)
∂FKt
∂qt−1( j) = γ
l,kt ( j) =
∂FKt
∂qt
∂qt
∂qt−1
=
(λ
l,kw,t ( j)+
L
∑l=1
λl,kc,t ( j)γ l
t+1( j)
)∂qt
∂qt−1(2.32)
The partial derivative of the current inflow respect to the previous inflow, from equation 2.28 is ex-
pressed as:
∂qt
∂qt−1= φt( j)
(σt( j)
σt−1( j)
)(2.33)
and λ kw,t (J×1), λ
h,khp,t (J×H×1) and λ
l,kc,t [L×1] are the dual information associated to water balance
(equation 2.13), the L cuts of the FBF (equation 2.12) and the H linear segments of the power functions
(equation 2.15), respectively.
Taking the expectation over the K artificially generated flows, the slope vectors ϕ lt and γ l
t,1 can be
determined:
ϕlt ( j) =
1K
K
∑k=1
ϕl,kt ( j) (2.34)
γlt ( j) =
1K
K
∑k=1
γl,kt ( j) (2.35)
Finally, the constant term is given by:
βlt =
1K
K
∑k=1
Fkt −∑
Jϕ
lt ( j)s◦t ( j)−∑
Jγ
lt,1( j)q◦t−1( j) (2.36)
As mention earlier, the backward optimization generates an outer approximation of the FBF Ft+1
(Figure 2.2). The accuracy of the approximation is evaluated at the end of the forward simulation. This
23
Sampled storage(period t, iteration 1)
Storage(period t+1, branch 1)
st + 1
Cut 1,1
Storage(period t+1, branch k)
Cut 1,k
Storage(period t+1, branch K)
Cut 1,K
Inflowbranch k
Ft + 1
AggregatedCut 1
Ft + 1
Ft + 1
Ft + 1
Sampled storage(period t, iteration 2)
Storage(period t+1, branch 1)
Cut 2,1
Storage(period t+1, branch k)
Cut 2,k
Storage(period t+1, branch K)
Cut 2,K
Inflowbranch k
Ft + 1
AggregatedCut 2
Ft + 1
Ft + 1
Ft + 1
Ft + 1
ApproximatedCut
Inflowbranch 1
Inflowbranch 1
Inflowbranch K
Inflowbranch K
st + 1
st + 1
st + 1
st + 1
st + 1
st + 1
st + 1
Figure 2.5: Backwards openings and the approximation of the FBF (adapted from Tilmant and Kelman(2007))
phase yields all the successive states and decision for each of M historical or synthetically generated
hydrologic sequences. Then the expected lower bound on the optimal solution is defined as:
Z =1M
T
∑t=1
bmt (st ,qm
t ,st+1,rt) =Zm
M(2.37)
where bmt is the immediate benefit at stage t for the hydrologic sequence m ∈ [1,2, . . . ,M]. This
forward simulation phase provides a lower bound with a 95% confidence intervals which allow us to
determine whether the upper bound is a good a approximation or not. If the upper bound does not fall
inside the confidence interval of the lower bound, the approximation is statistically not acceptable and
a new backward recursion is implemented with a new set of hyperplanes build on the storage volumes
that were visited during the last simulation phase.
24
The 95% confidence interval around the estimated lower bound Z is calculated as:[µZ−1.96
σZ√M,µZ +1.96
σZ√M
](2.38)
2.7 Climatological information as exogenous variable
Incorporating exogenous hydrologic variables into the state-space vector of SDDP offers the po-
tential to improve the performance of SDDP-derived release policies. The natural extension to in-
clude climatic variability into the autoregressive model is the MPAR model with exogenous variables
(MPARX). Using p previous inflows qt and b past exogenous variables Xt , the incremental flow at
node j, qt( j), can be derived from a multisite periodic autoregressive model with exogenous variables
MPARX(p,b):
qt( j)−µqt ( j)σqt ( j)
=p
∑i=1
φi,t( j)
(qt−i( j)−µqt−i( j)
σqt−i( j)
)+
b
∑κ=ι
ϑκ,t( j)
(Xt−κ( j)−µXt−κ
( j)σXt−κ
( j)
)+ εt( j)
(2.39)
where µXt and σXt are respectively the vectors of the periodic mean and the standard deviation of the
exogenous variables, and ϑκ,t the vector of the exogenous regressors. As indicated in Equation (2.39)
the exogenous variables may cover a different range of past input values, from ι to b, not necessarily
starting from t−1. This is significant in time-delay systems where the effect of an input may become
active after a certain time period (Marmarelis and Mitsis, 2014).
MPARX(p,b)
Exogenousvariables
Endogenousvariables.
𝑞𝑡−𝑝 𝑞𝑡
𝑋𝑡−𝜅
Figure 2.6: The ARX model structure (adapted from Ljung (1999))
2.8 Hydropower scheduling and contract management
Since the deregulation of the electricity market, in both developed and developing countries during
the eighties and nineties (Boubakri and Cosset, 1998), a variety of tools and methods have been
developed to jointly analyze operation scheduling and contract management (Mo et al., 2001; Gjelsvik
et al., 2010; Mo and Gjelsvik, 2002; Kristiansen, 2004; Flach et al., 2010). The typical hydro-based
producers face different types of risk such as price risk and quantity risk caused by both inflow and
25
demand uncertainty. In order to reduce the risk exposure, producers usually trade a variety of contracts
dealing with physical (power plants) and financial assets (energy contracts). As examples of contracts,
we can find future and forward contracts, option contracts, and load factor contract. The latter, usually
named as flexible contract is a physical or financial contract between two parts where price, energy and
maximum power (load factor) is predetermined, but its use is flexible, meaning that buyer determines
the amount of energy to be bought (Mo et al., 2001).
The mathematical description of a flexible contract follows the stage to stage energy balance, ac-
counting for the maximum possible withdrawal of energy ut = Pcτt [MWh], and the amount of energy
remaining of the contract wt [MWh]:
wt+1 = wt −ut (2.40)
with the initial energy amount of the flexible contract w0 [GWh]
wt = w0 (2.41)
and
Pwτt ≤ ut ≤ Pwτt (2.42)
where Pw [MW] is the instant power that can be withdrawn and τ the number of hours in period t.
Moreover, the demanded load Dt [MWh] must be met with the energy produced by the system Ptτ
and the sale/purchases through the contracts:
Ptτt +up,t −us,t = Dt (2.43)
These energy purchases up,t[MWh], and energy sales us,t[MWh] can be introduced in the immediate
benefit function (Equation 2.16) through the vector of slack variables zt , and consequently the prices
would be included in the vector ξt .
2.9 SDDPX formulation
A variant of the SDDP algorithm capable of incorporating various hydrologic information in the
decision-making process is developed. This new formulation, called SDDPX, incorporates exoge-
nous variables Xt , such as snow water equivalent and/or sea surface temperature in the state space
vector together with the previous inflows qt . Using p previous inflows qt and b past exogenous vari-
ables Xt , the vector of hydrologic state variables ht becomes [qt−1, qt−2,..., qt−p,Xt−κ ,..., Xt−b]. This
hydrologic information is encapsulated in SDDPX through a built-in MPARX model.
26
𝑠𝑡
𝑞𝑡 𝑟𝑡
𝑠𝑡+1Staget
𝑞𝑡+1 𝑟𝑡+1
𝑠𝑡+2Stage t+1
𝑠𝑇
𝑞𝑇 𝑟𝑇
𝑠𝑇+1StageT
𝑠1
𝑞1 𝑟1
𝑃1𝜏1
𝑠2Stage1
𝑤0 𝑤𝑡+1 𝑤𝑡+2 𝑤𝑇𝑤1 𝑤2
𝐷𝑡 𝐷𝑡+1 𝐷𝑇𝐷1
𝑢𝑝,1
𝑢𝑠,1
𝑃𝑡𝜏𝑡
𝑢𝑝,𝑡
𝑢𝑠,𝑡
𝑃𝑇𝜏𝑇
𝑢𝑝,𝑇
𝑢𝑠,𝑇
𝑃𝑡+1𝜏𝑡+1
𝑢𝑝,𝑡+1
𝑢𝑠,𝑡+1
𝑤𝑇+1
Figure 2.7: Schematization of energy trade
If at stage t, the system status is described by the storage st , the hydrological variable, and the amount
left of energy wt in C energy contracts, the one-stage SDDPX optimization problem can be written as:
Ft(st ,qt−1, ...,qt−p,Xt−κ , ...,Xt−b,wt) = max{bt(st ,qt ,rt ,st+1,ut)+Ft+1} (2.44)
subject to: Ft+1−ϕ
lt+1
ᵀst+1−χ
lt+1
ᵀwt+1 ≤ Γ
lt+1
ᵀht+1 +β
lt+1
...
Ft+1−ϕLt+1
ᵀst+1−χ
Lt+1
ᵀwt+1 ≤ Γ
Lt+1
ᵀht+1 +β
Lt+1
(2.45)
Γt+1ᵀht+1 = γt+1,1
ᵀqt + γt+1,2ᵀqt−1+, ...,γt+1,p
ᵀq(t−p)+1
γt+1,p+κᵀX(t−κ)+1+, ...,+γt+1,p+b
ᵀX(t−b)+1
(2.46)
the mass balance equation (2.13), the approximation of the hydropower functions 2.15, the trade
energy balance (2.43), the energy balance in the purchase contracts (2.40), and the lower and upper
boundaries (2.17 and 2.42).
In the backward phase the main modification to the traditional SDDP formulation lies in the calcula-
tion of the hyperplanes’ parameters ϕt+1, χt+1, βt+1, γt+1,1, γt+1,2,..., γt+1,p, γt+1,p+κ ,..., and γt+1,p+b
(see Equations 2.45 and 2.46). Using the Kuhn-Tucker conditions for optimality, the change of the
one-stage objective function Ft respect to the state variables st , wt , qt−1, qt−2,...,qt−p,Xt−κ ,..., and
27
Xt−b, can be determined by:
∂Ft
∂ st= λw,t +
H
∑h=1
λhhp,tψ
ht+1/2 (2.47)
∂Ft
∂wt= λe,t (2.48)
∂Ft
∂qt−1= λw,t
∂qt
∂qt−1+
L
∑l=1
λlc,t
(γ
lt+1,1
∂qt
∂qt−1+ γ
lt+1,2
)(2.49)
∂Ft
∂qt−2= λw,t
∂qt
∂qt−2+
L
∑l=1
λlc,t
(γ
lt+1,1
∂qt
∂qt−2+ γ
lt+1,3
)(2.50)
...∂Ft
∂qt−p= λw,t
∂qt
∂qt−p+
L
∑l=1
λlc,tγ
lt+1,1
∂qt
∂qt−p(2.51)
∂Ft
∂Xt−κ
= λw,t∂qt
∂Xt−κ
+L
∑l=1
λlc,t
(γ
lt+1,1
∂qt
∂Xt−κ
+ γlt+1,p+κ+1
)(2.52)
...∂Ft
∂qt−b= λw,t
∂qt
∂qt−b+
L
∑l=1
λlc,tγ
lt+1,1
∂qt
∂Xt−b(2.53)
where λw,t , λ hhp,t , λe,t and λ l
c,t are respectively the vectors with the dual information associated to the
mass balance (2.13), the H linear segments of the power functions (2.15), the energy balance (2.40)
and the L cuts of the benefit-to-go function (2.45).
Now, at stage t, s◦t , w◦t , q◦t−1, q◦t−2,..., q◦t−p, X◦t−κ ,..., and X◦t−b are sampled and K vectors of inflows qKt
are generated using the MPARX(p,b) (Equation 2.39). Since Fkt , which will be added to the expected
FBF at stage (t−1), can be approximated by:
Fkt ≤ ϕ
l,kt s◦t +χ
l,kt w◦t + γ
l,kt,1q◦t−1 + ...+ γ
l,kt,pq◦t−p + γ
l,kt,p+κX◦t−κ , ...,γ
l,kt,p+bX◦t−b +β
l,kt (2.54)
The slopes ϕl,kt and χ
l,kt are directly determined for each hydrologic scenario k using equations (2.47)
and (2.48):
ϕl,kt ( j) = λ
kw,t( j)+
H
∑h=1
λh,khp,t( j)ψh
t+1( j)/2 (2.55)
χl,kt (c) = λ
ke,t(c) (2.56)
and by using Equation 2.39 to find the partial derivative of qt respect to the hydrologic state variables,
the slopes γl,kt,1 , ...,γ
l,kt,p,γ
l,kt,p+κ , ..., and γ
l,kt,p+b can be determined using Equations (2.49) to (2.53) as:
28
γl,kt,1( j) = ϒt( j)
σqt ( j)σqt−1( j)
φt,1( j)+L
∑l=1
λl,kc,t ( j)γ l
t+1,2( j)
γl,kt,2( j) = ϒt( j)
σqt ( j)σqt−2( j)
φt,2( j)+L
∑l=1
λl,kc,t+1( j)γ l
t+1,3( j)
...
γl,kt,p( j) = ϒt( j)
σqt ( j)σqt−p( j)
φt,p( j)
...
γl,kt,p+κ( j) = ϒt( j)
σXt ( j)σXt−κ
( j)ϑt,κ( j)+
L
∑l=1
λl,kc,t+1( j)γ l
t+1,p+κ+1( j)
γl,kt,p+b( j) = ϒt( j)
σXt ( j)σXt−p( j)
ϑt,b( j) (2.57)
where ϒ( j) as:
ϒt( j) = λl,kw,t ( j)+
L
∑l=1
λl,kc,t ( j)γ l
t+1,1( j) (2.58)
Taking the expectation over the K artificially generated flows, the vector of slopes ϕ lt , χ l
t,1, γ lt,1,
γ lt,2,...,γ l
t,p, γ lt,p+κ ,...,γ l
t,p+b can be determined:
ϕlt ( j) =
1K
K
∑k=1
ϕl,kt ( j) (2.59)
χlt (c) =
1K
K
∑k=1
χl,kt (c) (2.60)
γlt,arx( j) =
1K
K
∑k=1
γl,kt,arx( j), (2.61)
∀arx = 1,2, ..., p, p+κ, ..., p+b
Finally, the constant term is given by:
βlt =
1K
K
∑k=1
Fkt −∑
Jϕ
lt ( j)s◦t ( j)−∑
Cχ
lt (c)w
◦t (c) . . .
−∑J
γlt,1( j)q◦t−1( j)−∑
Jγ
lt,2( j)q◦t−2( j)−∑
Jγ
lt,p( j)q◦t−p( j) . . .
−∑J
γlt,p+κ( j)X◦t−κ( j)−·· ·−∑
Jγ
lt,p+b( j)X◦t−b( j) (2.62)
As it was explain earlier, the accuracy of the FBF approximations is evaluated in a forward simulation
phase. Two different options exist to generate the M sequences required to simulate the system: i)
29
one can use the MPARX(p,b) to generate synthetic streamflow sequences based on historical records
of both endogenous and exogenous hydrologic variables, ii) or one can rather rely on series generated
outside of SDDPX using any relevant hydrologic model.
2.9.1 Toolbox description
The SDDPX model is coded in MATLAB®and the optimization scheme uses the linear Solver Gurobi®.
The toolbox processes two types of input data under Excel type format: i) the system configuration
file containing different characteristics of the physical river network, reservoir and demand data, and
some economic aspects (i.e. prices, contracts, energy demands), and ii) the hydrologic information
data file containing incremental flow data in m3/s for a set period of time: the incremental flow is
the natural runoff added to the network at each node. An additional Excel file is required for each
exogenous variable included in the analysis. These files must preserved the same structure of the
incremental flow data.
SDDPX Toolbox
Hydrologic Function MPARX model
Backward Optimization• FBF approximation
Forward Simulation• FBF Lower bound
Model Parametrization
Backwards openings (generated)
Forward sequences (historical endogenous and exogenous variables)
Re-optmization
Convergence criteria
Yes
no
Optimization modelSDDPX
Inputs
•Reservoir configuration•Topology•Physical and financial constraints
• Inflows• Exogenous variables
System configuration
Hydrologic information
Figure 2.8: Schematic SDDPX toolbox
The built-in MPARX model processes the hydrologic and climatologic information, determines the
parameters of the model and generates the backward openings required to implement the backward
optimization phase. Additionally, the Matlab function stores the historical inflows and the exogenous
variables, in such a manner to be used in the forward simulation and in the re-optimization phase. The
optimization module executes a main loop containing the backward and simulation phases. Once the
convergence criteria is met, a re-optimization procedure, in which the SDDPX-derived FBF approxi-
mations are used to determine the releases decision for the entire inflow series, is executed
30
Chapter 3
Case study
3.1 Outline
Two different hydropower systems have been used to test the different SDDP formulations: The
Gatineau River Basin and the Saguenay-Lac-Saint-Jean River Basin. This chapter describes the main
characteristics of the hydropower systems, the hydrologic regime, and the different management ob-
jectives.
3.2 The Gatineau River Basin and hydropower system
The Gatineau River, located in southwestern Quebec, rises in lakes north of the Baskatong Reservoir
and flows south to join the Ottawa River. The main river channel is about 400 km and drains an area of
23,700 km2. The watershed is used mainly for hydropower production and the cascade of power sta-
tions is operated by Hydro-Québec. Hydro-Québec is the main power generator in the province with
62 hydroelectric generating stations located across the province (installed capacity of 34,490 MW).
Hydro-Québec Production generates power for the Quebec market and sells its surpluses on whole-
sale markets primarily in Ontario and in the US. It is also active in arbitraging and purchase/resale
transactions (Retrieved from http://www.hydroquebec.com/generation/).
The Gatineau hydro-power system consists of a cascade of four power stations and two large reservoirs
(Figure 3.1). The upstream reservoir, Cabonga, has a storage capacity of 1.6 km3 and drains an area
of 2,201 km2. The second reservoir, Baskatong, was formed following the construction of the Mercier
Dam; it drains an area of 12,540 km2, and has a power plant with an installed capacity of 50 MW.
The Paugan station is a run-of-river power plant (R-O-R) with total capacity of 219 MW and a small
reservoir of 30 km2. The last R-O-R power stations are Chelsea and Rapides Farmer, with a total
installed capacity of 149 and 95 MW, respectively.
The watershed is characterized by a continental climate: warm and humid during the summer, and
31
Cabonga
Baskatong
Paugan
Chelsea
Inflow
Inflow
Inflow
Inflow
R. Farmers
Pembroke
Rapides Farmers Chelsea
Paugan
Baskatong
Cabonga
Joliette
Laval
MontrealOttawa
Cornwall
Longue
Trois-R
Reservoir
Power Plant
N
EW
S
77° W 76° W 75° W 74° W 73° W
48° N
47° N
46° N
45° N
Figure 3.1: Gatineau River Basin and hydropower system
generally wet, cold and snow covered in the winter. The nival regime can be described by (i) a
very high discharge during spring-summer (the snowmelt season; weeks 12-24), (ii) large variability
during spring and autumn and (iii) low flows during winter. Figure 3.2 displays the weekly averages
and weekly standard deviations of the inflows at four different nodes of the multireservoir system.
Table 3.1: Gatineau hydropower system
Storage Installed CapacityNode Name Hm3 MW
1 Cabonga 1633 -2 Baskatong 3175 503 Paugan 93 2194 Chelsea ROR 1485 Rapides Farmers ROR 95
a ROR: Run of the river power plant
A secondary objective of the operation of the multireservoir system is to control flooding at Wakefield.
32
4 8 12 16 20 24 28 32 36 40 44 48 520
10
20
30
40
50
60Node 1: Cabonga
[Hm
3]
4 8 12 16 20 24 28 32 36 40 44 48 520
50
100
150
200Node 3: Paugan
[Hm
3]
weeks
4 8 12 16 20 24 28 32 36 40 44 48 520
100
200
300
400
500Node 2: Baskatong
[Hm
3]
4 8 12 16 20 24 28 32 36 40 44 48 520
5
10
15
20
25
30Node 4: Chelsea
[Hm
3]
weeks
µq
t
σq
t
µq
t
σq
t
µq
t
σq
t
µq
t
σq
t
Figure 3.2: Weekly average (µ) and standard deviation (σ ) of inflow series - Gatineau River Basin
The town of Wakefield is located 35 km downstream of the Paugan plant and 20 km upstream of the
Chelsea plant. Wakefield is at the upstream end of the Chelsea reservoir. A rapid exists just upstream
of Wakefield, followed by a flatter river section. To prevent flooding at Wakefield, the Chelsea dam
pond is lowered by as much as 1 m at certain times. However, the lowering of the Chelsea pond
might not prevent minor floods at Wakefield, then the peaking capability at the Paugan plant is limited
(Abdelnour and Limited, 2001).
3.3 The Saguenay-Lac-Saint Jean River Basin and Rio Tinto system
Rio Tinto (RT), a major Aluminum producer in Canada, owns and operates a 3150MW-hydro-power
system in Quebec. The hydroelectric system, located in the Saguenay-Lac-Saint Jean River Basin,
includes 4 reservoirs and 5 hydropower plants: three on the Péribonca River and two on the Sague-
nay River. The drainage area is about 78,000 km2. In the northern part of the basin, there are two
large reservoirs: Manouane and Passes-Dangereuses with a storage capacity of 2.7 and 5.2 km3, re-
spectively. The downstream reservoir system, which includes the Lac-Saint-Jean Reservoir, drains
the Basins of the Péribonka, Ashuapmushuan, Mistassini and Mistassibi rivers. Figure 3.3 provides a
33
schematic representation of the RT hydroelectric network and Table 3.2 lists the main characteristics
of the system.
RLM
RPDCCP
CCS
SH
CIM SaguenayAlmaRLSJ
Riviere du Loup
Mis
tass
ibi
Riv
er
CCD
St Lawrence RivLEGEND
Power stationRiver basin
N
74⁰W 73⁰W
52⁰N
51⁰N
72⁰W 71⁰W 70⁰W
50⁰N
49⁰N
48⁰N
R. Lac-Manuane
R. Passes-Dangereuses
CCP
CCD
CCS
SH
CIM
Inflow
Inflow
Inflow
Inflow
Inflow
R. Chute-du-Diable
R. Lac-Saint-Jean
Figure 3.3: Rio Tinto hydropower system
Figure 3.4 displays the weekly averages and the standard deviations of inflows at 4 nodes of the
multireservoir system. We can see that the hydrologic regime is characterized by a high discharge
during spring-summer (weeks 14 to 24) and very low flows during winter (weeks 48 to 13); high
variability during spring influenced by the snowmelt season, and moderate variability during summer-
autumn influenced by the precipitation on the river basin. The total annual inflow at Lac-Manouane,
Passes-Dangereuses, Chute-du-Diable and Lac-Saint-Jean reservoirs are respectively 3.6; 7.3; 5.7 and
27.1 Km3. The total annual inflow at Chute-à-la-Savane is 0.924 Km3.
The water stored in Lac-Saint-Jean Reservoir includes direct rainfall onto the water body, natural
runoff, and water coming in from the Péribonka River, which is regulated by the Passes-Dangereuses
and Chute-du-Diable reservoirs. Lac-Saint Jean Reservoir is regulated by the Isle-Maligne (CIM)
34
Table 3.2: Rio Tinto hydro-power system characteristics
Storage Installed CapacityNode Id Name Hm3 MW
1 RLM Reservoir Lac-Manouane 2657 -2 RPD Reservoir Passes-Dangereuses 5228 7503 RCD Reservoir Chute-du-Diable 396 2054 CCS Chute-à-la-Savane ROR 2105 RLSJ Reservoir Lac-St-Jean 5083 4026 SH Shipshaw ROR 1587
a ROR: Run of the river power plant
4 8 12 16 20 24 28 32 36 40 44 48 520
50
100
150
200
250node:RLM
weeks
[Hm
3]
µqt
σqt
4 8 12 16 20 24 28 32 36 40 44 48 520
100
200
300
400
500
600node:RPD
weeks
[Hm
3]
µqt
σqt
4 8 12 16 20 24 28 32 36 40 44 48 520
100
200
300
400
500node:RCD
weeks
[Hm
3]
µqt
σqt
4 8 12 16 20 24 28 32 36 40 44 48 520
500
1000
1500
2000node:RLSJ
weeks
[Hm
3]
µqt
σqt
Figure 3.4: Weekly average and standard deviation of inflow series - Saguenay-Lac-St-Jean RiverBasin
35
power plant and the spillways on the Grande Décharge and Petite Décharge rivers, before flowing
towards the Saguenay River. During the spring flood period (i.e. April 1 to June 30), the volume of
water entering the Reservoir is over five times its storage capacity: spillways on the Petite Décharge
and the Grande Décharge tend to be open. During the summer, fluctuations in the reservoir levels are
influenced by precipitation, natural inputs, regulated inputs from the Péribonka River and hydroelec-
tric production at CIM (Retrieved from http://energie.riotinto.com, July 18 2017).
3.3.1 Recreational and fishing level constraints at LSJR
Reservoir Lac-St-jean has important recreational and sport-fishing industries which highly constrain
the storage levels during summer and autumn seasons. An agreement between the Government of
Québec and Rio Tinto, signed in 1986, established a maximum storage elevation of 101.84 m a.s.l.
However, since 1991 RT has voluntary established a maximum elevation of 101.54 m a.s.l. Fur-
thermore, it has been set that from July 24 to September 1, the minimum elevations is 100.78 m
a.s.l., and the maximum elevation must no exceed 101.39 m a.s.l. (Retrieved from http://energie.
riotinto.com, July 18 2017). Figure 3.5 displays a schematic diagram of the maximum and mini-
mum elevations at Lac-Saint-Jean Reservoir agreed between RT and the Government of Québec.
4 8 12 16 20 24 28 32 36 40 44 48 52Weeks
Ele
vation (
m a
.s.l.)
Min Level
Max Level
Historical Level
101.54
100.78
101.39
Figure 3.5: Schematic of minimum and maximum weekly levels at Lac-Saint Jean
3.3.2 Portfolio of energy contracts
RT hydro-power system has the capacity to produce more than 17 TWh/year, which constitutes
roughly 90% of the electricity required for the production of aluminum, forcing RT to buy energy
to fully ensure the production of the mineral. RT has two (2) energy contracts with Hydro-Quebec
limited in instant power and in total energy: P-Annual contract, available the entire year, and an 11-
week P-Pre-Freshet contract available after winter. Likewise, as part of the agreement, hydro-Quebec
purchases the RT power surpluses.
36
3.3.3 Climatological variables
Climate data, including precipitation, maximum and minimum temperature as well as snow water
equivalent (SWE), is collected by RTA’s private network of 22 weather stations and 11 SWE transect
field measurement sites. Figure 3.6 displays the weekly average SWE and weekly precipitation for
each node of the system.
4 8 12 16 20 24 28 32 36 40 44 48 520
150
300node:RLM
7SWE
<SWE
4 8 12 16 20 24 28 32 36 40 44 48 520
20
40node:RLM
4 8 12 16 20 24 28 32 36 40 44 48 520
150
300node:RPD
4 8 12 16 20 24 28 32 36 40 44 48 520
20
40node:RPD
4 8 12 16 20 24 28 32 36 40 44 48 52
Sno
w W
ater
Equ
ival
ent [
mm
]
0
150
300node:RCD
4 8 12 16 20 24 28 32 36 40 44 48 52Pre
cipi
tatio
n [m
m]
0
20
40node:RCD
4 8 12 16 20 24 28 32 36 40 44 48 520
75
150node:CS
4 8 12 16 20 24 28 32 36 40 44 48 520
20
40node:CS
weeks4 8 12 16 20 24 28 32 36 40 44 48 52
0
150
300node:RLSJ
weeks4 8 12 16 20 24 28 32 36 40 44 48 52
0
20
40node:RLSJ
7p
<p
Figure 3.6: Weekly average and standard deviation of Snow Water Equivalent and precipitation -Saguenay-Lac-St-Jean River Basin
37
Chapter 4
Overview of Results
4.1 Outline
As outlined in the introduction, this research focuses on the analysis of multireservoir system when
different hydrologic and climatic information are incorporated into the SDDP hydro-economic model.
Following the methodological activities presented therein, this chapter summarizes the main findings
of this research. More details are provided in the scientific Papers I and II.
4.2 Incorporation of multi-lag autocorrelation
In the first activity, an existing SDDP model with a built-in lag-1 hydrologic model, is modified to
take into account multiple lags through a MPAR(p) model. The hydrologic model developed during
this activity is capable of analyzing different spatially distributed hydrologic series, estimating the
parameters of the model, selecting the order p of the periodic model, and generating the two sets
of hydrologic information required to implement the SDDP algorithm. Annexe C presents the main
characteristics of the MPAR(p) model with an illustrative example of parameters estimation and order
selection.
The mathematical formulation to couple the higher order MPAR(p) hydrologic model and SDDP can
be found in Paper I. In that paper, we present an assessment of climate change impact on the operating
policies of the hydroelectric system in the Gatineau River Basin.
4.2.1 Results
Before exploring Paper I, this subsection presents the analysis of including the multi-lag autocorre-
lation in the decision making problem. A comparison between the lag-1 (SDDP(1)) and multi-lag
(SDDP(p)) is performed. Both SDDP models are implemented under the same conditions: K=25
backwards openings and M= 25 simulation sequences for a planing period of T = 260 weeks (5
years). The analysis is performed on the simulation results achieved after a re-optimization procedure
38
Figure 4.1: Draw-down refill cycle Baskatong and Cabonga Reservoirs. Lag-1 SDDP (Left panel) andmulti-lag SDDP(p) (right panel)
(Tejada-Guibert et al., 1993). The SDP-derived future benefit function approximations (Ft+1) from
the third year (t = 105,...,157), are used to determine the release decisions for the entire inflow series.
Figure 4.1 displays the draw-down refill cycle of Cabonga and Baskatong Reservoirs for the lag-1 and
the multi-lag SDDP models. Generally speaking, the annual cycles follow the same behavior: a draw-
down phase during the first weeks of the year (winter season) and a refill phase during the snow-melt
season (weeks 12-24, see hydrologic regime) . However, the storage levels reached using the multi-lag
model are much more variable than those achieved with the lag-1 SDDP. The information that can be
captured with the extended formulation better anticipates the incoming inflows and therefore yields
policies that better exploit the storage capacity of the reservoirs.
Figure 4.2 displays the statistical distribution of the annual spillage losses (panel (a)) and energy
production (panel (c)) for both formulations, SDDP(1) and SDDP(p). As we can see, regardless of the
non-exceedance probability the spillage losses are greater with the SDDP(1) model; panel (b) shows
the horizontal difference for the same probabilities. On average, the reduction spillage losses is around
824 Hm3/yr. Similarly, in panel (c) and (d), we can observe that 95% of the time the annual energy
production increases with the SDDP(p) model : an average annual increase of 78 GWh is identified
for the Gatineau hydro-power system.
39
0 5 10 150
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Spillage losses (Km3)
No
n-E
xce
ed
an
ce
pro
ba
bili
ty [-]
(a)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 1 2
Deviation (Km3)
(b)
2 2.5 3 3.5 40
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Energy (TWh)
(c)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
-0.2 0 0.2Deviation (TWh)
(d)
SDDP(1)
SDDP(p)
SDDP(1)
SDDP(p)
Figure 4.2: (a) Statistical distribution of annual spillage losses (b) spillage deviation respect SDDP(1)(c) Statistical distribution of annual energy production (d) energy production deviation with respectSDDP(1)
4.2.2 Operating policies facing climate change
In Paper I we develop and test a horizontal approach to assess climate change impact on water re-
sources systems, where hydro-climatic projections are not rapidly available. The typical approach
to assess climate change is based on a vertical integration/coupling of models. General Circulation
Models (GCM) provide future climatic projections, which are downscaled and used as inputs to hy-
drologic models. Afterwards, water systems models process these hydrologic projections to adapt the
operation policies to climate change. From the decision-making point of view this is challenging due
to the intensive computational and resource requirements, which are often increased when analyzing
more sources of uncertainty. Since hydro-climatic projections are becoming more often available for
different water systems, the purpose of the study is to analyze whether the computationally intensive
vertical approach is necessary in the first place, or the analysis could rely on projections available
in neighboring systems. To do so, we present a comparative assessment of the performance of the
Gatineau hydroelectric system under future climate conditions using the vertical approach and the
horizontal approach in which available projections are extrapolated from a neighboring system to the
system of interest.
Historical and projected climate information is provided by Hydro-Quebec, the Centre d’Expertise
Hydrique du Québec (CEHQ) and Ouranos. Climatic projections are obtained from the third gen-
eration coupled global climate model (CGCM3) of the Canadian Centre for Climate Modeling and
Analysis. Future projections are dynamically downscaled using the Canadian regional climate model
40
78° W 77° W
48° N
47° N
46° N
45° N
76° W 75° W 74° W
Cabonga
Baskatong
Paugan
Rapides Farmers ChelseaReservoir
Power Plant
Gatineau Basin
Du Lièvre Basin
G A T I N E A U
UD L I È V R E
N
EW
S
Ottawa
Cornwall
Pembroke
M
L
Figure 4.3: Gatineau and Du Lievre river basins
(CRCM version 4.2.3). Even though the uncertainty regarding the intensity of future consumption of
fossil fuels is carried out for various emission scenarios, our methodology is illustrated for scenario
A2 only.
Hydrologic projections are generated by CEHQ using the hydrologic model Hydrotel. To account for
the structural (hydrologic) uncertainty, inherent to the Gatineau River Basin, a multimodel approach
based on 20 different lumped hydrologic models is also used. Eventually a total of 21 time series
are made available for the stochastic multireservoir optimization model SDDP(p). Implementing
SDDP(p) on each time series yields an empirical distribution that reflects the impact of the structural
uncertainty of the hydrological processes on the energy generated by the system. This statistical
distribution constitutes the benchmark to assess the quality of the extrapolation techniques.
To implement the horizontal approach, the Du-Lievre River Basin which flows south and parallel to
the Gatineau River is selected as source river basin (Figure 4.3). The reference and future inflows
from DuLievre River Basin, as generated by the CEHQ using the Hydrotel model, are used to test four
different extrapolation techniques to derive the river discharges in the Gatineau River Basin. Table 4.1
summarizes the assumptions of the four extrapolation techniques tested. Further details in Paper I.
41
Table 4.1: Extrapolation techniques
Extrapolation technique AssumptionDrainage area ratio Both basins are hydrologically similar.Change in the mean Changes in the average weekly discharges in the
basin are proportional to those observed in theneighboring system.
Reference weekly regression Spatial correlation between river inflows remainsunchanged in the future. The current linear relation-ships will not be affected by climate change.
Autoregressive model The changes in mean and standard deviation areidentical to those observed in the source basin. Spa-tial correlation of the noise is preserved.
2 2.5 3 3.5 4Energy (TWh)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Non
-exc
eeda
nce
prob
abili
ty (
-)
Empirical CDF
Structural uncertaintyBenchmarkArea RatioChange in meanLinear RegressionAutoregressive Model
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
-20 0 20Relative difference (%)
Change in mean
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
-20 0 20
Area Ratio
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
-20 0 20
Linear Regression
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
-20 0 20
AutoregressiveModel
(a) (b)
Figure 4.4: Annual energy generation (a) cumulative distribution functions (b) relative differencesbetween the distribution functions
The SDDP(p) model of the Gatineau System is run 21 times, each time exploiting a different time
series of river discharges obtained from the 21 hydrologic models. The process was then repeated for
each extrapolation technique. As part of the results, Figure 4.4(a) presents the statistical distribution
of the annual energy generation using both the vertical and the horizontal approaches. The structural
uncertainty is represented in gray and its expected value (black CDF) is the benchmark against which
the extrapolation techniques are compared. This is the mean of the 21 different hydrologic models
that are part of the vertical approach directly implemented on the Gatineau River Basin.
Figure 4.4(b) displays the relative horizontal differences between the CDFs corresponding to the dif-
ferent extrapolation techniques and the benchmark CDF. These subplots show to what extent a par-
42
ticular extrapolation technique performs across contrasting hydrologic conditions. We can see that
the extrapolation based on the weekly linear regressions performs better than the others because the
relative differences are smaller over a large spectrum of probabilities. In other words, regardless of
the hydrologic conditions, the energy outputs obtained after extrapolating the flows with linear regres-
sions almost match those of the benchmark (with the exception of the driest year - low non-exceedance
probabilities).
The analysis reveals that the statistical distribution of energy generation can already be well ap-
proached using the extrapolation technique based on linear regressions. Further discussion on the
weekly basis can be found in Paper I. Generally speaking, annual and weekly productions are within
their respective structural uncertainty ranges, indicating that the extrapolation provides a reasonable
approximation of the amount of energy that can be expected.
For illustrative purposes, Figure 4.5 shows the expected draw-down refill cycle at Baskatong Reservoir
for the current hydrologic conditions and the hydrologic projections derived from the linear regres-
sion approach. We can identify that for the selected emission scenario, the refill phase starts earlier
following the earlier spring freshet. Since for this emission scenario, weekly discharges during the
snowmelt season are larger than under current conditions, the reservoir must be depleted earlier and a
further to free enough storage capacity, avoiding spillage losses during the snowmelt period.
0 4 8 12 16 20 24 28 32 36 40 44 48 52
500
1000
1500
2000
2500
3000
Sto
rage [hm
3]
Current Conditions
4 8 12 16 20 24 28 32 36 40 44 48 52
500
1000
1500
2000
2500
3000
Sto
rage [hm
3]
Expected storage value
Weeks
Current Condition
Climate change Horizontal app
Min Storage
Max Storge
0 4 8 12 16 20 24 28 32 36 40 44 48 52
500
1000
1500
2000
2500
3000
Sto
rage [hm
3]
Climate change - Horizontal approach
Figure 4.5: Baskatong drawdown-refill cycle - Climate change and current conditions
4.3 Incorporation of exogenous variables
Paper II presents SDDPX, an extension of SDDP capable of analyzing large scale water resources
allocation problems taking into account various hydro-climatic information. In practice, reservoir
inflows can be affected by other observable variables, which can help better capture the hydrologic
processes (e.g. snow water equivalent and/or sea surface temperature). Such variables, called exoge-
nous variables, are incorporated into the SDDP algorithm through a multi-site periodic autoregressive
model with exogenous variables (MPARX). With this extension the allocation policies are function of
the storage, the p previous inflows and relevant external variables. Paper II presents the new mathe-
matical formulation to derive the cuts approximating the future benefit functions based on the MPARX
43
model. The Rio Tinto (RT) system is used as case study to illustrate the potential gain associated with
the use of exogenous variables when operating a multireservoir system.
Four different formulation are tested (Table 4.2). The lag-1 SDDP(1), the multi-lag SDDP(p) and
two formulations based on the MPARX(p,b), differing in the selected exogenous variables. the snow
water equivalent SWEt and the combination of the SWEt and the accumulated winter precipitation Pwt
(i.e. accumulated precipitation from October until time t (Kim and Palmer, 1997)). A re-optimization
procedure is implemented in which SDDPX-derived future benefit functions are employed within a
simulation carried out over the historical record of both the endogenous and exogenous hydrologic
variables.
Table 4.2: SDDP formulations
MPARX(p,b)Formulation p b XSDDP(1) 1 - -SDDP(p) p - -SDDPX(p,SWE) p 1 SWESDDPX(p,SWE,Pw) p 1 SWE,Pw
4.3.1 Results
For this activity, the configuration is a simplified, hypothetical, system with a single objective: the
maximization of hydropower generation. Figure 4.6 illustrates the expected value of the aggregated
storage of the two largest reservoirs in the system: Passes-Dangereuses and Lac-Saint-Jean Reservoirs.
We can observe that the policies achieved using SDDPX show a more aggressive depletion of the
reservoirs during the draw-down phase (lower levels), indicating that the incorporation of exogenous
hydrologic variables can better anticipate the snow-melt runoff during spring. These lower levels
impact the levels beyond the summer therefore reducing the unproductive spills. In contrast, the
SDDP(1) and SDDP(p) tend to hedge more against hydrologic uncertainty, reaching higher levels
during the depletion phase in order to avoid potential water shortage in the future.
Table 4.3 summarizes a comparison of the different formulations with respect to SDDP(1). It shows
the mean spillage reduction, the mean release increase and the annual gain of energy. Basically, we can
see that increasing the number of lags from 1 to p already increases the amount of energy produced by
54 GWh. Incorporating exogenous variables further increases the production of energy by 46 GWh
and 51 GWh depending on whether SWE or the combination of SWE and winter precipitation are
used. Assuming an average market price of 45 US$/MWh (NYISO, 2017), these gains correspond to
a 2.45 to 4.70 million US$ increase in annual energy value. These energy gains are made possible by
improved operating policies that better exploit the storage capacity of the system.
Figure 4.7 presents the statistical distribution of the annual spillage losses (left panel) and the marginal
value of water (right panel) at each power plant for the SDDP(1) and SDDPX(p,SWE,Pw) formula-
tions. Generally speaking, results reveal that, regardless of the hydrologic condition, the spills are
44
4 8 12 16 20 24 28 32 36 40 44 48 522000
3000
4000
5000
6000
7000
8000
9000
10000
11000
Weeks
Sto
rag
e H
m3
SDDP(1)
SDDP(p)
SDDPX(p,SWE)
SDDPX(p,SWE,Pw)
Figure 4.6: Accumulated drawdown-refill cycle Passes Dangereuses and Lac-Saint-Jean reservoirs
Table 4.3: Average annual results - Differences with respect to the SDDP(1) model
Spillage Release Annual gainreduction increase of energy
Model (m3/s) (m3/s) (GWh)SDDP(p) 15 16 54SDDPX(p,SWE) 35 36 101SDDPX(p,SWE,Pw) 33 34 105
reduced when SDDPX is used to determine the policies; the horizontal gap between the lines repre-
sents the reduction of spillage at each power plant with respect to SDDP(1) formulation. Likewise,
the right panel shows the impact associated with the incorporation of more hydrological information
on the marginal value of water (i.e. Lagrange multiplier associated with the mass balance equation).
It reveals that at any given plant and regardless of the hydrologic condition the operator would be
willing to pay more for the same unit of water because the unproductive spills are reduced.
4.4 Joint optimization of physical and financial assets
In Paper II, a joint optimization of the physical infrastructure (i.e. Reservoirs and power plants) and
financial assets (i.e. energy contracts) is presented. As mentioned earlier, Rio Tinto’s hydropower
system is capable to produce 90% of the energy required to satisfy the production of aluminum,
45
35 40 45 50 550
0.5
1
Marginal water values $/1000 m3
RPD
18 20 22 24 260
0.5
1
Marginal water values $/1000 m3
RCD
14 16 18 200
0.5
1
Marginal water values $/1000 m3
Non-E
xceeda
nce pr
obabili
ty [-]
CS
10 11 12 13 140
0.5
1
Marginal water values $/1000 m3
RLSJ
7.5 8 8.5 9 9.50
0.5
1
Marginal water values $/1000 m3
SH
0 10 20 30 400
0.5
1
Spillage losses m3/s
RPD
0 20 40 600
0.5
1
Spillage losses m3/s
RCD
0 20 40 60 800
0.5
1
Spillage losses m3/sNon-E
xceeda
nce pr
obabili
ty [-]
CS
0 200 400 6000
0.5
1
Spillage losses m3/s
RLSJ
0 100 200 3000
0.5
1
Spillage losses m3/s
SH
SDDP(1) SDDPX(p,SWE,Pw)
Figure 4.7: Statistical distribution of the annual spillage losses (left panel) and the marginal value ofwater (right panel)
46
Table 4.4: Average annual results - Differences with respect to the SDDP(1) model
Annual gain Net Purchasesof energy reduction
Model (GWh) (%)SDDP(p) 30 4.0SDDPX(p,SWE) 50 6.6SDDPX(p,SWE,Pw) 51 6.7
forcing RT to buy energy to fully ensure the production of the mineral. RT has with Hydro-Québec
two (2) energy purchases contracts limited to a maximum annual amount of energy and maximum
instant power: P-Annual, available the entire year and P-Pre-Freshet only available 11 weeks during
the pre-freshet period. Besides, hydro-Québec has a power agreement with RT for the purchase of the
RT power surpluses. A fourth dummy and expensive contract (price 20 times higher than the regular
purchase price) is included in the configuration of the system to make sure that the load is met under
the worst conditions.
4.4.1 Results
This joint optimization also considers the secondary operating objectives such as recreation and flood
control at Lac-Saint-Jean. Results reveal that the expected gain in energy production using SDDP(p)
and SDDPX formulations is reduced. It means that when a more restrictive configuration is analyzed,
the operating policies tend to hedge more against the hydrologic risk, therefore accepting energy
deficits in the current period to reduce the probability of greater energy shortage in the future. This re-
duction in power output is the opportunity cost of hedging, and it is compensated by energy purchases.
Compared to SDDP(1) the net energy purchases are reduced by 6.6% and 6.7% when the exogenous
variables are included (Table 4.4). We can also see that the energy gains/purchases reduction seem
to follow the principle of diminishing marginal returns: the incorporation of the second exogenous
variable produce a marginal improvement.
Figure 4.8 displays the statistical distribution of the purchases made from the portfolio of contracts
using two different formulations: the simplest SDDP(1) and the most complex SDDPX formulation.
The use of contract "P-Pre-Freshet", which is available 20% of the year, is comparable for both for-
mulations. However, the SDDP(1) operating policy uses less energy from this ephemeral contract and
instead purchases more energy from the P-Annual and/or P-Extras contracts. The statistical distribu-
tion for the P-Extras contract reveals that regardless of the non-exceedance probability the amount
of energy bought is higher when the SDDP(1) formulation is used. These differences are marginal
if they are compared with those found in the P-annual contract statistical distribution. While with
the SDDP(1) formulation the maximum instant power is equaled 38% of the time (non-exceedance
probability > 0.62), with the SDDPX-derived policies this amount is equaled only 22% of the time.
Let us now analyze the annual power efficiency of the system (i.e. ratio between the power generated
47
0 25 50 75 1000
0.2
0.4
0.6
0.8
1
No
n-E
xce
ed
an
ce
pro
ba
bili
ty [
-]P-Pre-Freshet
0 25 50 75 1000
0.2
0.4
0.6
0.8
1
Weekly energy purchase [-]
P-Annual
0 25 50 75 1000
0.2
0.4
0.6
0.8
1P-Extras
SDDP(1)
SDDPX(p,SWE,Pw)
SDDP(1)
SDDPX(p,SWE,Pw)
SDDP(1)
SDDPX(p,SWE,Pw)
Figure 4.8: Statistical weekly distribution of energy purchases from the portfolio of contracts
-4 -2 0 2 4 60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
% Difference in Power Efficiency
No
n-E
xce
ed
an
ce
pro
ba
bili
ty [
-]
SDDP(1)
SDDPX(p,SWE,Pw) 1st config
SDDPX(p,SWE,Pw) 2nd config
Figure 4.9: % of difference in the power efficiency respect SDDP(1) formulation for both configura-tion
48
by the system and the total outflow). In general terms, the re-optimization results reveal a marginal im-
provement in the average power efficiency of the system when more hydrologic variables are included
in the state-vector: going from 1,399.7 to 1,401.4 W/m3/s when increasing the number of lags (from
1 to p), up to 1,403.3 and 1,403.4 W/m3/s when respectively (SWE) and (SWE,Pw) are incorporated
as exogenous variables. In order to compare the results obtained from the two different configuration
(i.e. maximization of hydropower generation and joint optimization with secondary operating ob-
jectives), Figure 4.9 displays the statistical distribution of the annual differences in power efficiency
between SDDPX(p,SWE,Pw) and SDDP(1) formulation. Two main findings can be observed. First,
for both configurations, the efficiency of the system is enhanced more than 65 % of the time (i.e. non-
exceedance probability greater than 0.35) when the exogenous hydrologic variables are incorporated
in the analysis. Secondly, the variability of the efficiency gains/losses is less pronounced for the sec-
ond configuration. In that case, both the upside and the downside of the hydrological risk associated
with imperfect forecasts are partly suppressed. As energy shortages are increasingly costly, i.e. com-
pensated by purchases through increasingly more expensive contracts, release policies for the second
configuration act as a hedging mechanism thereby yielding more conservative release decisions over
the entire spectrum of the hydrologic uncertainty.
49
Conclusion
As the availability of various hydro-climatic information keeps increasing due to new technologies,
such as high resolution satellite imagery, there is a need to incorporate such information into the
decision making processes that we use to manage our water resources. Up until recently, the main
bottleneck was to be found in the optimization algorithms which were unable to handle either a large
quantity of hydro-climatic information or large water resource systems. This trade-off is now largely
removed by new algorithmic strategies which seek to identify approximate solutions to the complex
decision-making problem. One of such algorithms is Stochastic Dual Dynamic Programming (SDDP).
This PhD thesis fits within this context and aims at assessing the economic value of various hydro-
logic and climatic information when managing a hydropower system consisting of physical (reser-
voirs and power stations) and financial assets (contracts). To achieve this, the methodology relies on
a gradual improvement of an available SDDP model that uses the past inflows as the only hydrologic
state variable. The extended formulation jointly optimizes the production of hydroelectricity and the
sale/purchase of energy through multiple contracts taking into account various hydro-climatic data.
The original SDDP model was first upgraded to consider the multi-lag autocorrelation in the stream-
flow series. The built-in MPAR(1) was substituted by a MPAR(p) model and the SDDP formulation
was adjusted accordingly. The resulting SDDP(p) was then tested on a hydroelectric system in the
Gatineau River Basin in Quebec. The simulation results showed that the SDDP(p) -derived policies
perform better than SDDP(1), especially during the snowmelt season. It turns out that the inclusion of
multiple lags improves the forecast of the incoming flows, therefore avoiding unproductive spills at the
end of the refill phase. This SDDP(p) formulation was also used to determine the optimal operating
strategies of the Gatineau hydropower system under climate change.
The second extension, denoted SDDPX, incorporates various hydro-climatic data such as snowpack,
precipitation or sea surface temperature, through the use of exogenous hydrologic state variables.
Here, at each time step, a built-in MPAX(p,b) model processes both past inflows (lag p) and past
exogenous data (lag b). The parameters of the MPARX model are then used to approximate the
objective function to be maximized by the SDDPX algorithm. The implementation of SDDPX is
illustrated with the hydro-power system of Rio Tinto (RT) located in the Saguenay-Lac-Saint-Jean
River Basin. A hypothetical system configuration, which aimed maximizing hydropower generation
only, revealed the extent of the energy gain that can be expected when more hydrologic information
50
is incorporated into SDDP. Since the hydrologic processes are better captured and the uncertainty
reduced, SDDPX-derived policies are even more wiling to take risk than SDDP(p) policies. The
joint optimization of physical and financial assets showed that all four SDDP-derived polices hedge
more against potential costly shortages, inducing a reduction in power output. Nevertheless, SDDPX-
derived policies still revealed better performances in terms of net purchases reductions, meaning that
the opportunity cost of hedging is lower since the hydrologic risks is reduced when more hydrologic
information is included in the analysis. However, the gains associated with the incorporation of this
information seems to follow the principle of diminishing marginal returns. Therefore, an input variable
selection assessment would be necessary to determine the number of variables to be incorporated.
Although the optimization model was tested for two hydropower systems characterized by a snow-
driven hydrology, SDDPX is capable of analyzing large-scale water systems for different flow regimes,
and consequently process relevant exogenous variables. The current SDDPX forward phase is carried
out based on historical records of both, exogenous and endogenous hydrologic variables, which might
become a limitation if limited information is available.
51
Future work
The optimization model SDDPX developed in this research should be used as departing point for
future research. The future research may be focused on, but not limited, the following specific recom-
mendations.
• Evaluate the economic value of using the SDDPX-derived FBF approximations as terminal
value function in the short-term optimization. Author suggests the incorporation of ensemble
streamflow predictions ESP to determine the SDDPX terminal value functions, and assess their
potential to enhance the short-term operating policies.
• Test the SDDPX model in water systems with a different hydrologic regime, other than the
snow-melt regime analyzed in this project. Since macro-climatic phenomenon signals such El
Niño Southern Oscillation (ENSO) and Pacific Decadal Oscillation PDO are now well recog-
nized for forecasting streamflow in some regions, author suggests the implementation of SD-
DPX model in large-scale system where these climatic signals have demonstrated the potential
of enhancing the streamflow foresight.
• Implement SDDPX on water systems with other water use sectors, such irrigated agriculture,
municipal water supply, navigation, environmental flows, etc, and assess the economic impact
of the climatic variables on the water allocation in these different sectors.
• Incorporate the stochasticity of the exogenous variables in the SDDPX-FBF approximations,
using MIMO-type hydrologic models (i.e. Multiple Input Multiple Output). While the back-
ward optimization is solved for k synthetic openings, the variability on the exogenous variables
was omitted, as they are not required in the mass balance.
• Further analysis in the sampling strategies of the hydrologic state are required. Preliminary
analysis have shown an important variability in the performance of the system.
52
Bibliography
Abdelnour, R., and F. T. Limited (2001), Ice Booms in Rivers ; Lessons Learned and the Development
of Reliable Solutions Ice Booms Main Characteristics The Wakefield Ice Boom, 11th Work. River
Ice, pp. 1–10.
Ahmad, A., A. El-Shafie, S. F. Mohd Razali, and Z. S. Mohamad (2014), Reservoir optimization in
water resources: A review, doi:10.1007/s11269-014-0700-5.
Anghileri, D., N. Voisin, A. Castelletti, F. Pianosi, B. Nijssen, and D. P. Lettenmaier (2016), Value
of long-term streamflow forecasts to reservoir operations for water supply in snow-dominated river
catchments, Water Resour. Res., 52(6), 4209–4225, doi:10.1002/2015WR017864.
Archibald, T., K. McKinnon, and I. Thomas (1997), An aggregate stochastic dynamic programming
model of multireservoir systems, Water Resour. Res., 33(2), 333–340.
Archibald, T., K. McKinnon, and I. Thomas (2006), Modeling the operation of multireservoir systems
using decomposition and stochastic dynamic programming, Nav. Res. Logist., 53(3), 217–225, doi:
10.1002/nav.
Beale, E. (1955), On Minimizing a convex function subject to Linear Inequalities, J. R. Stat. Soc. Ser.
B, 17, 173–184.
Bellman, R. (1957), Dynamic Programming, 1 ed., Princeton University Press, Princeton, NJ, USA.
Bellman, R., and S. Dreyfus (1962), Applied dynamic programming, Tech. rep., United States Air
force project RAND, Princeton University Press.
Boubakri, N., and J.-C. Cosset (1998), The financial and operating performance of newly privatized
firms: Evidence from developing countries, J. Finance, 53(3), 1081–1110, doi:10.1111/0022-1082.
00044.
Box, G.E.P. and Jenkins, G.M. (1970). Time Series Analysis Forecasting and Control. Holden Day,
San Francisco.
Bras, R. L., R. Buchanan, and K. C. Curry (1983), Real time adaptive closed loop control of
reservoirs with the High Aswan Dam as a case study, Water Resour. Res., 19(1), 33–52, doi:
10.1029/WR019i001p00033.
53
Carpentier, P. L., M. Gendreau, and F. Bastin (2013), Long-term management of a hydroelectric mul-
tireservoir system under uncertainty using the progressive hedging algorithm, Water Resour. Res.,
49(5), 2812–2827, doi:10.1002/wrcr.20254.
Cerisola, S., J. M. Latorre, and A. Ramos (2012), Stochastic dual dynamic programming applied to
nonconvex hydrothermal models, Eur. J. Oper. Res., 218(3), 687–697, doi:10.1016/j.ejor.2011.11.
040.
Côté, P., D. Haguma, R. Leconte, and S. Krau (2011), Stochastic optimisation of Hydro Quebec
hydropower installations: a statistical comparison between SDP and SSDP methods, Can. J. Civ.
Eng., 1434(April), 1427–1434, doi:10.1139/l11-101.
Dantzig, G. (1955), Linear programming under Uncertainty, Manag. Sci. 1, pp. 197–206.
Day, G. N. (1985), Extended Streamflow Forecasting Using NWSRFS, J. Water Resour. Plan. Manag.,
111(2), 157–170, doi:10.1061/(ASCE)0733-9496(1985)111:2(157).
De Queiroz, A. R., and D. P. Morton (2013), Sharing cuts under aggregated forecasts when decom-
posing multi-stage stochastic programs, Oper. Res. Lett., 41(3), 311–316, doi:10.1016/j.orl.2013.
03.003.
Desreumaux, Q., P. Côté, and R. Leconte (2014), Role of hydrologic information in stochastic dy-
namic programming: a case study of the Kemano hydropower system in British Columbia, Can. J.
Civ. Eng., 41(9), 839–844, doi:10.1139/cjce-2013-0370.
Faber, B., and J. Stedinger (2001), Reservoir optimization using sampling SDP with ensemble stream-
flow prediction (ESP) forecasts, J. Hydrol., 249, 1113–133, doi:10.1016/S0022-1694(01)00419-X.
Flach, B., L. Barroso, and M. Pereira (2010), Long-term optimal allocation of hydro generation for a
price-maker company in a competitive market: latest developments and a stochastic dual dynamic
programming approach, IET Gener. Transm. Distrib., 4(2), 299, doi:10.1049/iet-gtd.2009.0107.
Garen, D. C. (1992), Improved Techniques in Regression-Based Streamflow Volume Forecasting, J.
Water Resour. Plan. Manag., 118(6), 654–670, doi:10.1061/(ASCE)0733-9496(1992)118:6(654).
Gelati, E., H. Madsen, and D. Rosbjerg (2010), Markov-switching model for nonstationary runoff con-
ditioned on El Niño information, Water Resour. Res., 46, W02,517, doi:10.1029/2009WR007736.
Gelati, E., H. Madsen, and D. Rosbjerg (2014), Reservoir operation using El Niño forecasts case study
of Daule Peripa and Baba, Ecuador, Hydrol. Sci. J., 59(8), 1559–1581, doi:10.1080/02626667.2013.
831978.
Georgakakos, A. (1989), The Value of Streamflow Forecasting in Reservoir Operation, Water Resour.
Bull., 25(4).
54
Gjelsvik, A., B. Mo, and A. Haugstad (2010), Long and Medium term Operations Planning and
Stochastic Modelling in Hydro-dominated Power Systems Based on Stochastic Dual Dynamic Pro-
gramming, Hanb. power Syst. I, Energy Syst., pp. 33–55, doi:10.1007/978-3-642-02493-1.
Goor, Q., R. Kelman, and A. Tilmant (2011), Optimal multipurpose multireservoir operation model
with variable productivity of hydropower plants, J. water Resour. Plan. Manag., 137(3), 258–267,
doi:10.1061/(ASCE)WR.1943-5452.
Hannan, E. J., and M. Deistler (2012), The Statistical Theory of Linear Systems, Society for Industrial
and Applied Mathematics, doi:10.1137/1.9781611972191.
Hart, W. E., C. Laird, J.-P. Watson, and D. L. Woodruff (2012), Pyomo – Optimization Modeling in
Python, vol. 67, 13–28 pp., doi:10.1007/978-1-4614-3226-5.
Homen-de Mello, T., V. L. de Matos, and E. C. Finardi (2011), Sampling strategies and stopping cri-
teria for stochastic dual dynamic programming: a case study in long term hydrothermal scheduling,
Energy Syst., 2(1), 1–31, doi:10.1007/s12667-011-0024-y.
Infanger, G. (1993), Planning under Uncertainty: Solving large scale stochastic linear programs,
Boyd and Fraser publishing company, Danvers, Massachusetts.
Infanger, G., and D. P. Morton (1996), Cut sharing for multistage stochastic linear programs with
interstage dependency, Math. Program., 75(2), 241–256, doi:10.1007/BF02592154.
Johnson, S. A., J. R. Stedinger, C. A. Shoemaker, Y. Li, and J. Tejada-Guibert (1993), Numerical
solution of continous state dynamic programs using linear and spline interpolation, Oper. Res.,
41(3), 484–500.
Karamouz, M., and H. V. Vasiliadis (1992), Bayesian stochastic optimization of reservoir operation
using uncertain forecasts, Water Resour. Res., 28(5), 1221–1232, doi:10.1029/92WR00103.
Karamouz, M., F. Szidarovszky, and B. Zahraie (2003), Water Resources Systems Analysis with Em-
phasis on Conflict Resolution, 608 pp.
Karunanithi, N., W. J. Grenney, D. Whitley, and K. Bovee (1994), Neural Networks for River Flow
Prediction, J. Comput. Civ. Eng., 8(2), 201–220, doi:10.1061/(ASCE)0887-3801(1994)8:2(201).
Kelman, J., J. Stedinger, L. Cooper, E. Hsu, and S. Yuan (1990), Sampling stochastic dynamic
programming applied to reservoir operation, Water Resour. Res., 26(3), 447–454, doi:10.1029/
WR026i003p00447.
Kim, Y., and R. Palmer (1997), Value of seasonal flow forecasts in Bayesian stochastic programming,
J. water Resour. Plan. Manag., 123(6), 327–335, doi:10.1061/(ASCE)0733-9496(1997)123:6(327).
55
Kim, Y.-o., H. Eum, E. Lee, and I. Ko (2007), Optimizing operational policies of a Korean multireser-
voir system using sampling stochastic dynamic programming with ensemble streamflow prediction,
J. water Resour. Plan. Manag., 133(1), 4–14.
Kitanidis, P. K., E. F. I. Foufoula-georgiou, and S. Anthony (1987), Error analysis of conventional
discrete and gradient dynamic programming, Water Resour. Res., 23(5), 845–858.
Kristiansen, T. (2004), Financial risk management in the electric power industry using stochastic
optimization, Adv. Model. Optim., 6(2), 17–24.
Marmarelis, V., and G. Mitsis (Eds.) (2014), Data-driven Modeling for Diabetes, Lecture Notes in
Bioengineering, Springer Berlin Heidelberg, Berlin, Heidelberg, doi:10.1007/978-3-642-54464-4.
Labadie, J. W. (2004), Optimal Operation of Multireservoir Systems: State-of-the-Art Review, J.
Water Resour. Plan. Manag., 130(2), 93–111, doi:10.1061/(ASCE)0733-9496(2004)130:2(93).
Ljung, L. (1999), System Identification: Theory for the User, in Wiley Encycl. Electr. Electron. Eng.,
John Wiley & Sons, Inc., Hoboken, NJ, USA, doi:10.1002/047134608X.W1046.
Ltkepohl, H. (2007), New Introduction to Multiple Time Series Analysis, Springer Publishing Com-
pany, Incorporated.
Lohmann, T., A. S. Hering, and S. Rebennack (2015), Spatio-temporal hydro forecasting of mul-
tireservoir inflows for hydro-thermal scheduling, Eur. J. Oper. Res., 255(1), 243–258, doi:10.1016/
j.ejor.2016.05.011.
Macian-Sorribes, H., A. Tilmant, and M. Pulido-Velazquez (2016), Improving operating policies of
large-scale surface-groundwater systems through stochastic programming, Water Resour. Res., doi:
10.1002/2016WR019573.
Mo, B., A. Gjelsvik, and A. Grundt (2001), Integrated risk management of hydro power scheduling
and contract management, Power Syst. IEEE Trans. , 16(2), 216–221, doi:10.1109/59.918289.
Mo, B., and A. Gjelsvik (2002), Simultaneous optimisation of withdrawal from a flexible contract and
financial hedging, 14th Power Syst. Comput. Conf., (June), 24–28.
NYISO (2017), New York Independent System Operator.
Oliveira, R., and D. P. Loucks (1997), Operating rules for multireservoir systems, Water Resour. Res.,
33(4), 839–852, doi:10.1029/96WR03745.
Pagano, T. C., D. Garen, and S. Sorooshian (2004), Evaluation of Official Western U. S. Sea-
sonal Water Supply Outlooks , 1922 – 2002, J. Hydrometeorol., 5(5), 896–909, doi:10.1175/
1525-7541(2004)005<0896:EOOWUS>2.0.CO;2.
56
Pereira, M. (1989), Optimal stochastic operations scheduling of large hydroelectric systems, Int. J.
Electr. Power Energy Syst., 11(3), 161–169, doi:10.1016/0142-0615(89)90025-2.
Pereira, M., N. Campodónico, and R. Kelman (1998), Long-term Hydro Scheduling based on Stochas-
tic Models, Int. Conf. Electr. Power Syst. Oper. Manag., pp. 1–22.
Pereira, M. V. F., and L. M. V. G. Pinto (1991), Multi-stage stochastic optimization applied to energy
planning, Math. Program., 52(1-3), 359–375, doi:10.1007/BF01582895.
Pereira-Cardenal, S. J., B. Mo, A. Gjelsvik, N. D. Riegels, K. Arnbjerg-Nielsen, and P. Bauer-
Gottwein (2016), Joint optimization of regional water-power systems, Adv. Water Resour., 92, 200–
207, doi:10.1016/j.advwatres.2016.04.004.
Pina, J., A. Tilmant, and F. Anctil (2016), Horizontal Approach to Assess the Impact of Climate
Change on Water Resources Systems, J. Water Resour. Plan. Manag., p. 04016081, doi:10.1061/
(ASCE)WR.1943-5452.0000737.
Pina, J., A. Tilmant, and P. Côté (2017), Optimizing Multireservoir System Operating Policies Using
Exogenous Hydrologic Variables, Water Resour. Res., (1990), 1–20, doi:10.1002/2017WR021701.
Poorsepahy-Samian, H., V. Espanmanesh, and B. Zahraie (2016), Improved Inflow Modeling in
Stochastic Dual Dynamic Programming, J. Water Resour. Plan. Manag., p. 04016065, doi:
10.1061/(ASCE)WR.1943-5452.0000713.
Pritchard, G. (2015), Stochastic inflow modeling for hydropower scheduling problems, Eur. J. Oper.
Res., 246, 496–504, doi:10.1016/j.ejor.2015.05.022.
Rani, D., and M. M. Moreira (2009), Simulation Optimization Modeling: A Survey and Potential
Application in Reservoir Systems Operation, Water Resour. Manag., 24(6), 1107–1138, doi:10.
1007/s11269-009-9488-0.
Raso, L., P. O. Malaterre, and J. C. Bader (2017), Effective Streamflow Process Modeling for Optimal
Reservoir Operation Using Stochastic Dual Dynamic Programming, J. Water Resour. Plan. Manag.,
p. 04017003, doi:10.1061/(ASCE)WR.1943-5452.0000746.
Rockafellar, R. T., and R. J. B. Wets (1991), Scenarios and Policy Agregation in Optimization under
Uncertainty, Math. Oper. Res., 16(1), 119–147, doi:10.1287/moor.16.1.119.
Rotting, T., and A. Gjelsvik (1992), Stochastic dual dynamic programming for seasonal scheduling in
the Norwegian power system, IEEE Trans. Power Syst., 7(1), 273–279, doi:10.1109/59.141714.
Saad, M., P. Bigras, a. Turgeon, and R. Duquette (1996), Fuzzy Learning Decomposition for the
Scheduling of Hydroelectric Power Systems, Water Resour. Res., 32(1), 179–186, doi:10.1029/
95WR02971.
57
Salas, J., Delleur, J., Yevjevich, V., and Lane, W. (1980). Applied Modeling of Hydrologic Time Series.
Water Resources Publications, Littleton, Colorado, USA.
Sene, K. (2010), Hydrometeorology, 1–427 pp., Springer Netherlands, Dordrecht, doi:10.1007/
978-90-481-3403-8.
Séguin, S., S.-E. Fleten, P. Côté, A. Pichler, and C. Audet (2017), Stochastic short-term hydropower
planning with inflow scenario trees, Eur. J. Oper. Res., 259(3), 1156–1168, doi:10.1016/j.ejor.2016.
11.028.
Shapiro, A., W. Tekaya, J. P. da Costa, and M. P. Soares (2013), Risk neutral and risk averse Stochastic
Dual Dynamic Programming method, Eur. J. Oper. Res., 224(2), 375–391, doi:10.1016/j.ejor.2012.
08.022.
Stedinger, J. R., B. F. Sule, and D. P. Loucks (1984), Stochastic dynamic programming mod-
els for reservoir operation optimization, Water Resour. Res., 20(11), 1499–1505, doi:10.1029/
WR020i011p01499.
Tejada-Guibert, J., S. Johnson, and J. Stedinger (1993), Comparison of two approaches for imple-
menting multireservoir operating policies derived using stochastic dynamic programming, Water
Resour. Res., 29(12), 3969–3980.
Tejada-Guibert, J. A., S. A. Johnson, and J. R. Stedinger (1995), The Value of Hydrologic Informa-
tion in Stochastic Dynamic Programming Models of a Multireservoir System, Water Resour. Res.,
31(10), 2571–2579, doi:10.1029/95WR02172.
Thomas, H.A. and Fiering, M.B. (1962). Mathematical synthesis of streamflow sequences for the
analysis of river basins by simulation. In Design of Water Resource Systems, Edited by Mass et al.,
Harvard University Press, Cambridge, 1962.
Tilmant, A., and R. Kelman (2007), A stochastic approach to analyze trade-offs and risks as-
sociated with large-scale water resources systems, Water Resour. Res., 43(6), W06,425, doi:
10.1029/2006WR005094.
Tilmant, A., and W. Kinzelbach (2012), The cost of noncooperation in international river basins, Water
Resour. Res., 48(1), W01,503, doi:10.1029/2011WR011034.
Tilmant, A., D. Pinte, and Q. Goor (2008), Assessing marginal water values in multipurpose
multireservoir systems via stochastic programming, Water Resour. Res., 44(12), W12,431, doi:
10.1029/2008WR007024.
Todini, E. (1988), Rainfall-runoff modeling - Past, present and future, J. Hydrol., 100(1-3), 341–352,
doi:10.1016/0022-1694(88)90191-6.
58
Turgeon, A., and R. Charbonneau (1998), An aggregation disaggregation approach to long term reser-
voir management, Water Resour. Res., 34(12), 3585–3594.
Yeh, W. W.-G. (1985), Reservoir Management and Operations Models: A State-of-the-Art Review,
Water Resour. Res., 21(12), 1797–1818, doi:10.1029/WR021i012p01797.
Zahraie, B., and M. Karamouz (2004), Hydropower Reservoirs Operation : A Time Decomposition
Approach, Sci. Iran., 11(April), 92–103.
59
Appendix A
Paper I: Horizontal Approach to assesthe Impact of Climate Change on WaterResources Systems
Résumé
L’approche généralement utilisée pour évaluer les impacts du changement climatique sur des sys-
tèmes de ressources en eau est fondée sur une intégration et/ou un couplage vertical de modèles. Étant
donné que la gamme d’incertitude qui peut être explorée avec le GCM est limitée, les chercheurs
préfèrent s’appuyer sur des ensembles pour élargir la propagation, ce qui rend l’approche par modéli-
sation encore plus exigeante en termes de temps de calcul et de ressources. Lorsqu’un système d’eau
particulier doit être analysé, la question est de savoir si cette approche verticale à forte intensité de
calcul est nécessaire en premier lieu, ou s’il est possible de compter sur des projections disponibles
dans les systèmes voisins ? L’étude proposée aborde cette question en comparant les performances
d’un système de ressources en eau dans des conditions climatiques futures, en utilisant les approches
verticales et horizontales. La méthodologie est illustrée par le système hydroélectrique du bassin de la
rivière Gatineau au Québec (Canada). L’analyse des résultats de la simulation, obtenus à partir des ap-
proches verticales et horizontales, révèle que les productions énergétiques annuelles et hebdomadaires
du système sont similaires.
http://ascelibrary.org/doi/abs/10.1061/(ASCE)WR.1943-5452.0000737.
60
Horizontal Approach to Assess the Impact of ClimateChange on Water Resources Systems
Jasson Pina1; Amaury Tilmant, Ph.D.2; and Francois Anctil, Ph.D.3
Abstract: The typical approach to assess climate change impacts on water resources systems is based on a vertical integration/coupling of
models. Since the range of uncertainty that can be explored with GCM is limited, researchers rely on ensembles to enlarge the spread, making
the modeling approach even more demanding in terms of computation time and resource. When a particular water system must be analyzed,
the question is to know whether this computationally intensive vertical approach is necessary in the first place or if we could rely on pro-
jections available in neighboring systems? The proposed study addresses this question by comparing the performance of a water resource
system under future climate conditions using the vertical and the horizontal approaches. The methodology is illustrated with the hydropower
system of the Gatineau River Basin in Quebec, Canada. The analysis of the simulation results reveals that the annual and weekly energy
productions of the system derived from both the vertical and the horizontal approaches are similar. DOI: 10.1061/(ASCE)WR.1943-
5452.0000737. © 2016 American Society of Civil Engineers.
Author keywords: Water resources systems; Climate change; Hydropower generation; Stochastic programming; Multimodel hydrology.
Introduction
The profligate use of fossil fuel energy that was initiated by the
industrial revolution in the nineteenth century has had profound
impacts on societies and on the environment worldwide. The mas-
sive release of carbon dioxide in the atmosphere by ever increasing
human activities is expected to change the composition of the
atmosphere and, hence, the components of the hydrological cycle
such as precipitation and temperature (IPCC 2013). Significant ef-
forts have been devoted over the past 30 years to assess the impacts
of climate change on water resources and the effectiveness of
adaptation measures to mitigate those impacts.There are essentially two categories of adaptation measures to
climate change in the water sector: structural and nonstructaural.
The first one includes traditional engineering-based solutions such
as investments in hydraulic infrastructure like levees and dams
(e.g., Jeuland and Whittington 2014). In the second category, alter-
native solutions like insurance mechanisms (e.g., Botzen and van
der Bergh 2008), demand management (e.g., Iglesias and Garrote
2015), and changes in operating policies (e.g., Vicuna et al. 2009)
have received a lot of attention over the last decade.The typical approach to assess climate change impacts on water
resources systems is based on a vertical integration/coupling of
models: GCM models are run to project future precipitations
and temperatures, which are then downscaled and used as inputsto hydrologic models whose outputs are processed by watersystems models (Yao and Georgakakos 2001; Carpenter andGeorgakakos 2001; Harrison and Whittington 2002). From a deci-sion-making point of view, this top-down vertical approachpresents some challenges. Since the range of uncertainty that canbe explored with GCM is limited, researchers are relying on ensem-bles to enlarge the spread, making the modeling approach evenmore demanding in terms of computation time and resource. Forexample, in an attempt to cover all sources of uncertainty, Chenet al. (2011) combine various emission scenarios, initial GCM con-ditions, GCMs, downscaling techniques, and hydrological models.
Another challenge associated with this vertical, top-down ap-proach lies in the difficulty of attaching probabilities to the scenar-ios and hence to carry out a risk analysis. Yet, designing a portfolioof adaptation measures requires that the effectiveness of these mea-sures, whether they are structural or nonstructural, be properly as-sessed. Decision analysis is the theoretical framework that hastraditionally been implemented to deal with this planning issue.When rooted in a traditional benefit-cost analysis framework, de-cision analysis requires that key uncertainties be well characterized,which is often not the case with climate change scenarios (Dessaiand Hulme 2004; New et al. 2007).
To address this issue, Wilby and Dessai (2010), Prudhommeet al. (2010), and Brown et al. (2012) propose an alternative ap-proach for climate risk assessment that links bottom-up, stochasticvulnerability analysis with the traditional top-down use of GCMprojections. This alternative approach, coined Decision Scaling,has been used to manage climate risks in the Upper Great Lakesby Brown et al. (2011). Decision scaling is a robust decision mak-ing approach, a set of decision analytic methods to help decisionmakers identify solutions that satisfactorily perform across a widerange of potential climate changes (Lempert and Schlesinger 2000).
These two approaches (top-down versus bottom-up) illustratethe ongoing debate among scientists on how best to tackle the chal-lenges associated with climate risk assessment and management.On the one hand, proponents of the top-down approach argue theneed to improve the predictive capability of climate models in orderto reduce the uncertainty attached to climate projections. On the
1Ph.D. Student, Dept. of Civil Engineering and Water Engineering,Université Laval, Av. de la Médecine 1065 Pavillon Adrien-Pouliot,2983, QC, Canada G1V 0A6. E-mail: [email protected]
2Professor, Dept. of Civil Engineering and Water Engineering, Univer-sité Laval, Av. de la Médecine 1065 Pavillon Adrien-Pouliot, 2978, QC,Canada G1V 0A6 (corresponding author). E-mail: [email protected]
3Professor, Dept. of Civil Engineering and Water Engineering, Univer-sité Laval, Av. de la Médecine 1065 Pavillon Adrien-Pouliot, 2988, QC,Canada G1V 0A6. E-mail: [email protected]
Note. This manuscript was submitted on October 31, 2015; approved onSeptember 8, 2016; published online on November 10, 2016. Discussionperiod open until April 10, 2017; separate discussions must be submittedfor individual papers. This paper is part of the Journal of Water Resources
Planning and Management, © ASCE, ISSN 0733-9496.
© ASCE 04016081-1 J. Water Resour. Plann. Manage.
J. Water Resour. Plann. Manage., 04016081
Dow
nloa
ded
from
asc
elib
rary
.org
by
UN
IVE
RSI
TE
LA
VA
L o
n 11
/14/
16. C
opyr
ight
ASC
E. F
or p
erso
nal u
se o
nly;
all
righ
ts r
eser
ved.
61
other hand, proponents of the bottom-up approach rather considerthat, in light of the deep uncertainties involved, more efforts mustbe devoted to understanding the climate-related vulnerabilities(Weaver et al. 2013).
For managers of large hydropower systems, this scientific debateis important, but so are more practical (pressing) issues. One of themis how to rapidly appraise the impacts of a changing flow regime onthe energy produced by a particular system. Because the verticalapproach is computationally intensive, and with the increasingavailability of projections in neighboring systems, the question isto know whether we could obtain a decent approximation of thechanges in the energy generated by a particular system by extrapo-lating the hydro-climatic projections already available in a broaderregion encompassing that system. This study addresses this questionby comparing the performance of a water resource system underfuture climate conditions using the vertical approach and the hori-zontal approach in which available projections are extrapolated tothe system of interest. The methodology is illustrated with a cascadeof power stations on the Gatineau River in Quebec, Canada.
The paper is organized as follows: section “Material andMethod” presents the modelling of the climate and the hydrologicalprocesses, the multireservoir operation model, and then the selectedextrapolation methods. Section “Analysis of Simulation Results”describes the case study, which is followed by a presentationand a discussion of simulated results (energy generation). Finally,conclusions are given in section “Conclusion.”
Material and Method
This section starts with an overview of the methodology, which willbe explain in more detail in the following subsections. This meth-odology, which is illustrated on Fig. 1, is implemented on the hy-dropower system of the Gatineau River Basin in Quebec, Canada.This system consists of a cascade of four power stations and twolarge reservoirs located in the Laurentian region north of Ottawa.The top-down approach implemented in many river basins in theregion relies on the projections of the Canadian GCM, dynamicallydownscaled, and processed by a hydrological model to projectfuture discharges. Hydrologic projections are then used as inputsto a stochastic multireservoir optimization model (stochastic dualdynamic programming, SDDP) to derive energy outputs from thesystem. To assess the horizontal approach, we need a benchmark:the hydrologic projections derived from an extended vertical ap-proach implemented in the Gatineau River Basin (Fig. 1). Theextension is needed to account for the structural (hydrologic)uncertainty inherent to that basin. To achieve this, the vertical ap-proach in the Gatineau also relies on another set of 20 hydrologicmodels, each having a different structure. In a neighboring riverbasin (Du Lievre), the vertically obtained projections are extrapo-lated to the Gatineau River Basin using four different extrapolationtechniques. A reservoir operation model then processes all hydro-logic projections. Energy simulation results, derived either directlyfrom the vertical approach or indirectly by extrapolating the pro-jections available in Du Lievre River Basin, are compared.
To summarize, comparing the performance of the system underthe vertical and the horizontal approaches requires• H hydrologic models to capture the structural uncertainty in the
target river basin (here the Gatineau). These models are part ofthe vertical approach implemented in that basin. The mean ofthe projections derived from these models is used as a bench-mark to compare the accuracy of the extrapolation techniques.Here, H ¼ 21.
• K hydrologic sequences representative of the natural uncertaintyin the target river basin. These sequences are required by the
SDDP algorithm to derive the optimal reservoir operatingpolicies. Here, K ¼ 23.
• E extrapolation techniques are used to derive river dischargesin the target river basin (Gatineau) based on the projectionsavailable in the source river basin (Du Lievre). Here, fourextrapolation techniques (E ¼ 4) are tested and compared interms of energy generation and reservoir trajectories.The following subsections describe the methodology in more
detail.
Gatineau River Basin and Hydropower System
The Gatineau River, located in southwestern Quebec, rises in lakesnorth of the Baskatong Reservoir and flows south to join theOttawa River. The main river channel is about 400 km and drainsan area of 23,700 km2. The hydroelectric system consists of a cas-cade of four power stations and two large reservoirs (Fig. 2). Theupstream reservoir, Cabonga, has a storage capacity of 1.633 km2
and drains an area of 2,201 km2. The second reservoir, Baskatong,was formed following the construction of the Mercier Damdrains an area of 12,540 km2, and has a power plant with an in-stalled capacity of 55 MW. The Paugan station is a run-of-riverpower plant (R-O-R) with total capacity of 219 MW and a smallreservoir of 30 km2. The last R-O-R power stations are Chelseaand Rapides Farmer, with a total installed capacity of 149 and95 MW, respectively.
Climate Modeling and Projections
Historical data were provided by Hydro-Québec and the Centred’Expertise Hydrique du Québec (CEHQ), and climatic scenarios
Fig. 1. Schematics of the methodology
© ASCE 04016081-2 J. Water Resour. Plann. Manage.
J. Water Resour. Plann. Manage., 04016081
Dow
nloa
ded
from
asc
elib
rary
.org
by
UN
IVE
RSI
TE
LA
VA
L o
n 11
/14/
16. C
opyr
ight
ASC
E. F
or p
erso
nal u
se o
nly;
all
righ
ts r
eser
ved.
62
by Ouranos. Climatic projections were obtained from the third-generation coupled global climate model (CGCM3) of theCanadian Centre for Climate Modelling and Analysis, whichhas a surface grid with a roughly spatial resolution of 3.75°latitude/longitude; daily data for several 2D variables are availablefor years 2001–2100. Future projections were dynamically down-scaled to match the temporal and spatial scales of the hydrologicalmodels, using the Canadian regional climate model (CRCMversion 4.2.3), time slice simulation for 2041–2070 over the NorthAmerican domain. The CRCM horizontal grid is uniform in a polarstereographic projection, with a typical 45-km grid mesh (true at60°N), and its vertical resolution is variable using a Gal-Chenscaled height terrain-following coordinate (Music and Caya 2007).Local Intensity scaling method were implemented by Ouranos tocorrect biases between observations and simulations for the refer-ence period (1961–1990).
To account for the uncertainty regarding the intensity of futureconsumption of fossil fuels, the analysis should be carried out forvarious emission scenarios. However, the proposed methodologywill be illustrated for scenario A2 only.
Hydrologic Modeling and Projections
Hydrologic projections were generated by CEHQ using the hydro-logical model Hydrotel, a semi-physically based distributed modellargely used in Quebec for dam management (Fortin et al. 2001;Turcotte et al. 2004). To account for the structural (hydrologic)
uncertainty inherent to the Gatineau River Basin, a multimodel ap-proach based on 20 different lumped hydrologic models was alsoimplemented. The model selection was initially carried out byPerrin (2000) and revised by Seiller et al. (2012) for hydrologicalprojection purposes. It is mainly based on known performance andstructural diversity: from 4 to 10 free parameters, and 2 to 7 storageunits. The structural diversity ensures that the simple average of themodels’ output is much more robust to temporal transposabilitythan the output of the models taken individually (Seiller et al.2012). This arises from the fact that the models are somehow un-correlated and tend to cancel out each other’s errors. The gain thuslies in the combination of models and not in their individual per-formance. Indeed, Thiboult and Anctil (2015) compared Hydrotelto the above multimodel in validation and for short-term forecastingover 38 basins spread across the province of Quebec. Results invalidation showed that Hydrotel clearly stands among the bestmodels for all catchments, but is outperformed by the simple aver-age of the 20 lumped (much simpler) models. As for Thiboult andAnctil (2015), all models were applied here in exactly the sameconditions: they were run on a daily time step and fed with identicalprecipitations and potential evapotranspirations (computed fromOudin et al. (2005) temperature-based formulation). Snow accumu-lation and melt are simulated with the CemaNeige snow accountingmodule (Valéry et al. 2014).
Split sampling test (SST) (Klemeš 1986) was implemented forthe calibration and validation procedures. The principle is to cal-ibrate the models on a time series with selected characteristics and
Cabonga
Baskatong
Paugan
Chelsea
Inflow
Inflow
Inflow
Inflow
R. Farmers
Pembroke
Rapides Farmers Chelsea
Paugan
Baskatong
Cabonga
Joliette
Laval
MontrealOttawa
Cornwall
Longue
Trois-R
Reservoir
Power Plant
N
EW
S
77° W 76° W 75° W 74° W 73° W
48° N
47° N
46° N
45° N
Fig. 2. Gatineau River Basin [data from CEHQ 2012; Esri, DeLorme, TomTom, Intermap, Increment P Corp, GEBCO, USGS, FAO, NPS,
NRCAN, GeoBase, IGN, Kadaster NL, Ordnance Survey, Esri Japan, METI, Esri China (HongKong), Swisstopo, MapmyIndia, © OpenStreetMap
Contributors, and the GIS User Community (2015); basemaps ArcGIS]
© ASCE 04016081-3 J. Water Resour. Plann. Manage.
J. Water Resour. Plann. Manage., 04016081
Dow
nloa
ded
from
asc
elib
rary
.org
by
UN
IVE
RSI
TE
LA
VA
L o
n 11
/14/
16. C
opyr
ight
ASC
E. F
or p
erso
nal u
se o
nly;
all
righ
ts r
eser
ved.
63
to validate it on a contrasted time series. Model calibration wasperformed for the period 1969–1988, and the validation phase from1988 to 2005. The shuffled complex evolution (SCE-UA) (Duanet al. 1993) was used to calibrate the model parameters. The ob-jective function was the root-mean square error applied to theroot-squared transformed streamflow (RMSEsqrt). The evaluationof model performance in validation relied on the Nash-Sutcliffe efficiency criterion (Nash and Sutcliffe 1970) calculatedon the root-squared transformed streamflows.
Once all hydrological models were calibrated, climate projec-tions were used to obtain time series of future inflows at each nodeof the system (reservoirs and ROR power plants) as pictured onFig. 2. A total of 21 time series were made available for the sto-chastic multireservoir optimization model: one set from the distrib-uted hydrologic model (Hydrotel) and 20 sets from the lumpedhydrologic models. The 21 time series capture the structural uncer-tainty associated with the hydrologic processes in the river system(here the Gatineau). This uncertainty is inherent to the system andwould be present regardless of the emission scenario considered. Inother words, for a given set of meteorological inputs and at a givenpoint in space and time, an empirical distribution of river flows canbe traced out from the simulations of the 21 hydrologic models.Implementing the multireservoir optimization model on each timeseries yields an empirical distribution of the energy generated atthat point, i.e., at a particular power plant and time of the year. Thisstatistical distribution reflects the impact of the structural uncer-tainty of the hydrological processes on the energy generated bythe system. It will be used to assess the quality of the extrapolationtechniques discussed in the next section.
Extrapolating Future River Discharges
All daily hydrologic times were aggregated on a weekly timestep in order to be compatible with the reservoir operation model.Different extrapolation techniques were used to derive the riverdischarges in the Gatineau from the weekly projections availablein neighbouring systems (river basins). The reference and futureinflows from Du Lievre River Basin, as generated by the CEHQusing Hydrotel, were selected to test different extrapolation tech-niques. Du Lievre River Basin drains an area of 9,946 km2, flowssouth and parallel to the Gatineau River, and empties into theOttawa River (Fig. 3). Four different techniques were tested1. Drainage area ratio: The first, most simple, extrapolation tech-
nique relies on the ratios between the drainage areas. Future pro-jection at Du Lievre River Basin were directly extrapolated tothe Gatineau River Basin by using Eq. (1) and the drainagearea ratios presented in Table 1. This technique makes the as-sumption that both river basins, Gatineau and Du Lievre, arehydrologically similar
QFutðjÞ ¼ QFutð6ÞAðjÞAð6Þ ð1Þ
where j ¼ 1, : : : ,5 = subbasins in the Gatineau River System;j ¼ 6 = Du Lievre River Basin. QfutðjÞ are the future inflowprojections and and AðjÞ is the (sub)basin area.
2. Change in mean: The second technique extrapolates flows in theGatineau Basin from the ratios between future and current aver-age weekly flows in Du Lievre River Basin. Here, we assumethat changes in the average weekly discharges in one basin areproportional to those observed in the neighbouring system.Based on this assumption, the inflows in the Gatineau sub-basins are given by
QFut;tðjÞ ¼ QRef;tðjÞμFut;tð6ÞμRef;tð6Þ
ð2Þ
where μRef;t and μFut;t = average weekly discharges for thereference series and the future projections, respectively.
3. Reference weekly regression: The reference weekly regressionapproach exploits the linear relationship among the referenceinflows in both rivers basins, QrefðjÞ
QRef;tðjÞ ¼ atðjÞt · QRef;tð6Þ þ ctðjÞ ð3ÞParameters aðjÞ and cðjÞ are used to determine the future
values in the Gatineau River sub-basins using the hydrologicprojections in the Du Lievre River QFut;tð6Þ
QFut;tðjÞ ¼ atðjÞ · QFut;tð6Þ þ ctðjÞ ð4ÞThis third extrapolation technique assumes that current linear
relationships between weekly river discharges in both basinswill not be affected by climate change. In other words, for any
78° W 77° W
48° N
47° N
46° N
45° N
76° W 75° W 74° W
Cabonga
Baskatong
Paugan
Rapides Farmers ChelseaReservoir
Power Plant
Gatineau Basin
Du Lièvre Basin
G A T I N E A U
UD L I È V R E
N
EW
S
Ottawa
Cornwall
Pembroke
M
L
Fig. 3. Gatineau and Du Lievre River Basins [data from CEHQ 2012;
Esri, DeLorme, TomTom, Intermap, Increment P Corp, GEBCO,
USGS, FAO, NPS, NRCAN, GeoBase, IGN, Kadaster NL, Ordnance
Survey, Esri Japan, METI, Esri China (HongKong), Swisstopo, Map-
myIndia, © OpenStreetMap Contributors, and the GIS User Commu-
nity (2015); basemaps ArcGIS]
Table 1. Drainage Area Ratios
(j) Basin Area (km2) Ratio
1 Cabonga 2,201 0.222 Baskatong 12,540 1.263 Paugan 7,740 0.784 Chelsea 1,189 0.125 Rapides Farmers — —
6 Du Lievre 9,946 1
© ASCE 04016081-4 J. Water Resour. Plann. Manage.
J. Water Resour. Plann. Manage., 04016081
Dow
nloa
ded
from
asc
elib
rary
.org
by
UN
IVE
RSI
TE
LA
VA
L o
n 11
/14/
16. C
opyr
ight
ASC
E. F
or p
erso
nal u
se o
nly;
all
righ
ts r
eser
ved.
64
given week, the spatial correlation between river flows in bothbasins remains unchanged in the future regardless of the extentof the changes in river flows.
4. Correlation approach: The last technique takes into account thechanges in the weekly means and standard deviations fromDu Lievre River while assuming that the spatial correlationamong the inflows in the reference period remains identical.This extrapolation technique was chosen because it has thesame structure as the built-in hydrological model found in themutireservoir stochastic optimization model described inthe next section.The keep the notation simple, let us assume that the inflows
can be described by a lag-1 multisite autoregressive model (seeAppendix for a general formulation to p lags)
�
QtðjÞ − μtðjÞσtðjÞ
�
Ref
¼ ϕt
�
Qt−1ðjÞ − μt−1ðjÞσt−1ðjÞ
�
Ref
þ ϵtðjÞ ð5Þ
where σt = weekly standard deviation; and ϵtðjÞ = time independent(but spatially correlated) stochastic noise with mean zero andweekly variance σ2
��ε;t. ϕt is the lag-1 autoregressive parameterand is determined by using the method of the moments as thecorrelation ρi;t. Likewise, the variance is determined as σ2
��ε;t ¼1 − ϕ2
t . Once the parameters for the reference series are determined,the changes in the mean, standard deviation and in the autoregres-sive parameters from the reference inflows to the future projectionscalculated in Du Lievre River Basin
Δμt ¼μFut;tð6ÞμRef;tð6Þ
; Δσt ¼σFut;tð6ÞσRef;tð6Þ
; Δϕt ¼ϕFut;tð6ÞϕRef;tð6Þ
ð6Þ
are used to project the parameters for the future period on theGatineau River
μFut;tðjÞ ¼ Δμt · μRef;tðjÞσFut;tðjÞ ¼ Δσt · σRef;tðjÞϕFut;tðjÞ ¼ Δϕt · ϕRef;tðjÞ ð7Þ
Assuming that the noise ϵt follows a three-parameters lognormal distribution with μv, σv and lower bound ψt
fϵt ¼1
ðϵt − ψtÞffiffiffiffiffiffiffiffiffiffi
2πσv
p e−0.5f½logðϵt−ψtÞ−μv�=σvg2 ð8Þ
with mean μ��εt and variance σ2��ε;t
μ��εtψt þ eðμvþσ2v=2Þ σ2
��ε;te2ðμvþσ2vÞ þ eð2μvþσ2vÞ ð9Þ
Then the weekly future projections in the Gatineau sub-basinsare synthetically generated from time t0 to period T (number ofyears times number of weeks). The starting points of the generationprocedure Q0
t−1 are retrieved form the reference period and pro-jected directly using Eq. (13)
ψtðjÞ ¼ −μtðjÞσtðjÞ
− ϕt
�
Q0t−1ðjÞ − μt−1ðjÞ
σt−1ðjÞ
�
Fut
ð10Þ
λtðjÞ ¼ 1þ σ2ϵ;t
ψ2t ðjÞ
μvðjÞ ¼ 0.5Logσ2ϵ;t
λðjÞ½λðjÞ − 1� ð11Þ
σvðjÞ ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
log½λðjÞ�p
ð12Þ
QFut;tðjÞ ¼ σFut;t · eVtðjÞ·σvðjÞþμvðjÞ ð13Þ
where Vt = cross-correlated white noise.
Stochastic Multireservoir Optimization Model—SDDP
Reservoir operation problems are multistage, stochastic and non-linear optimization problem. The goal is to determine a sequenceof optimal decisions xt that maximize the expected sum of benefitsfrom system operation Z, over a planning period time T, whilemeeting operational and/or institutional constraints (Fig. 4).
The mathematical formulation of the multistage decision-making problem can be written as
maxxt
fZg ¼ maxxt
�
E
�
X
T
t¼1
αtbtðst; qt; rtÞ þ αTþ1νðsTþ1; qTÞ��
ð14Þ
where btð·Þ = immediate benefit function at stage t; νð·Þ = terminalvalue function; αt = discount factor at stage t; E½·� = expectationoperator; st = vector of storages at the beginning of stage t;qt = vector of inflows; and rt = vector of releases. The immediatebenefit function will include the net benefits from system genera-tion and penalties for not meeting target demands and/or violatingconstraints.
The immediate benefit function, at time t, can be written as
btðst; qt; rtÞ ¼ τ tðπ − θÞ 0Pt − ξ 0t zt ð15Þ
where τ t = number of hours in period t; Pt (MW) = vector of powergenerated; π = vector of energy price ($=MWh); θ = vector of theoperation and maintenance cost ($=MWh); and zt = vector ofdeficits or surpluses (unit deficit or surplus) penalized by the vectorξ 0t of penalties ($=unit deficit or surplus).
The stage to stage transformation function corresponds to themass balance equation
stþ1 − CRðrt þ ltÞ ¼ st þ qt − et ð16Þ
where stþ1 = vector of storage at the end of the period; lt = vector ofspills; CR = reservoir system connectivity matrix, CR
j;k ¼ 1ð−1Þwhen reservoir j receives (releases) water from (to) reservoirk; and et = vector of evaporation losses.
This problem can be solved with discrete stochastic dynamicprogramming (SDP), which recursively constructs a benefit-to-go function Ftþ1 at the grid points defining the state space. Withtwo reservoir storages and a minimum of four hydrologic state var-iables, the reservoir operation problem in the Gatineau cannot besolved with SDP as the computational effort increases exponen-tially with the number of state variables. To deal with the dimen-sionality issue, we use stochastic dual dynamic programming, an
Fig. 4. Multistage decision-making problem
© ASCE 04016081-5 J. Water Resour. Plann. Manage.
J. Water Resour. Plann. Manage., 04016081
Dow
nloa
ded
from
asc
elib
rary
.org
by
UN
IVE
RSI
TE
LA
VA
L o
n 11
/14/
16. C
opyr
ight
ASC
E. F
or p
erso
nal u
se o
nly;
all
righ
ts r
eser
ved.
65
extension of SDP that can handle a much larger state space (Tilmantet al. 2008).
The solution approach is based on the approximation of the ex-pected cost-to-go functions of SDP by piecewise linear functions(hyperplanes), meaning there is no need to evaluate Ftþ1 over adense grid. SDDP uses an iterative procedure to increase the accu-racy of the solution by adding new hyperplanes in the regions ofthe state-space where they are most needed, i.e., in the regions ex-plored by the system.
To further save computation time, SDDP relies on an analyticalrepresentation of the natural hydrologic uncertainty through amulti-site periodic autoregressive model MPAR(p). More detailson the MPAR(p) model are provided in Appendix. The natural un-certainty must not be confused with the structural uncertainty dis-cussed earlier. The natural uncertainty originates from thestochasticity of the hydrologic processes generating runoff. Thestructural uncertainty comes from our inability to perfectly modelthese processes. To describe the natural uncertainty, several realiza-tions of the stochastic processes are considered through, forexample, multiyear time series of river flows, while the structuraluncertainty is captured through the use of multiple hydrologicmodels, each having a different structure.
As mentioned earlier, SDDP is an iterative algorithm organizedaround two phases: a backward optimization and a forward simu-lation. In the backward phase, the MPAR model is used to generateK inflows scenarios (backward openings) at each node of thesystem. These scenarios are needed to analytically calculate the hy-perplanes’ parameters, and ultimately to derive the upper bound tothe true expected benefit-to-go function. The accuracy of thisapproximation is evaluated at the end of a forward simulation phasebased on M hydrological sequences representative of the naturaluncertainty. If the upper bound is inside the confidence interval ofthe lower bound, the approximation is statistically acceptable andthe algorithm stops, otherwise a new backward recursion is imple-mented with a new set of hyperplanes build on the storage volumesthat were visited during the last simulation phase.
Analysis of Simulation Results
The SDDP model of the Gatineau System was run 21 times, eachtime exploiting a different time series of river discharges obtainedby one of the 21 hydrologic models. The process was then repeatedfor each extrapolation technique discussed above. For each run, theoptimal operating policies were simulated over 23 historical hydro-logic sequences of 260 weeks (M ¼ 23 and T ¼ 260 weeks). Twas chosen so as to be sufficiently long to avoid the effectsof the boundary conditions on reservoir operating policies on in-termediate years (Goor et al. 2011). Here, results are analyzedfor year three only as the first and last two years are influencedby the boundary conditions (initial storages at time t ¼ 1 as wellas FTþ1 ¼ 0).
Table 2 lists the average and standard deviation (Std) of annualenergy generation in the system after extrapolating the hydrologicprojections in Du Lievre River Basin using the four techniques de-scribed above. The standard deviation reflects the natural uncer-tainty of the hydrologic processes in the basin. For example, forthe extrapolation technique based on the drainage areas, the annualhydroelectric production could vary from 2.5 to 3.8 TWh depend-ing on whether we have a (naturally) dry or a wet year. On average,depending on the extrapolation technique, the annual productionswould vary from 3.13 to 3.29 TWh. This 5% difference is signifi-cant, yet within the range of the structural uncertainty representedby the grey area on Fig. 5.
The mean of the multimodel (black CDF on Fig. 5) is the bench-mark against which the extrapolation techniques must be com-pared. Recall that this is the mean of the 21 different hydrologicmodels that a part of the vertical approach directly implementedon the Gatineau River Basin. We are interested in the horizontaldifferences between the CDFs derived after extrapolating the pro-jections and the benchmark; the smaller the differences across awide range of hydrologic conditions (nonexceedance probabilities),the better is the extrapolation technique. Fig. 6 displays the hori-zontal differences between the CDFs corresponding to the differentextrapolation techniques and the mean of the CDFs correspondingto all 21 hydrologic models (the benchmark). This figure showsto what extent a particular extrapolation technique performsacross contrasting hydrologic conditions, from very dry years(low nonexceedance probabilities) to very wet years (high nonex-ceedance probabilities). The examination of Fig. 6 reveals that theextrapolation based on the weekly linear regressions performsbetter than the others because the relative differences are smallerover a large spectrum of probabilities. In other words, regardlessof the hydrologic conditions, the energy outputs obtained afterextrapolating the flows with linear regressions almost match thoseof the benchmark (with the exception of the driest year).Table 3 shows the root-mean-square differences (TWh) betweenthe energy outputs obtained after extrapolating the flows and thebenchmark for various probability ranges (100, 90, 80, and50%) centered on the median. As expected from the visual exami-nation of Fig. 6, we can see that the extrapolation approach basedon weekly linear regressions generates the smallest differences withthe benchmark, especially for close to a normal hydrologic condi-tions (50% probability range). However, the predictive capability ofthe linear regression degrades as more extreme hydrologic years(wet or dry) are included in the analysis. Note that the latter is truefor all extrapolation techniques.
Table 2. Annual Energy Generation: Mean and Standard Deviation
EnergyArearatio
Changein mean
Linearregression AR model
Mean (TWh) 3.29 3.13 3.22 3.17Std (TWh) 0.3190 0.3545 0.3007 3.1916
2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 40
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Energy (TWh)
Non
exce
eden
ce p
roba
bilit
y (
)
Empirical CDF
Structural uncertainty
Benchmark
Area Ratio
Change in mean
Linear Regression
Autoregressive Model
Fig. 5. Annual energy generation—cumulative distribution functions
© ASCE 04016081-6 J. Water Resour. Plann. Manage.
J. Water Resour. Plann. Manage., 04016081
Dow
nloa
ded
from
asc
elib
rary
.org
by
UN
IVE
RSI
TE
LA
VA
L o
n 11
/14/
16. C
opyr
ight
ASC
E. F
or p
erso
nal u
se o
nly;
all
righ
ts r
eser
ved.
66
As the technique based on the weekly linear regression performsbetter for this case study, it seems that climate change will not affectthe spatial correlation between weekly river discharges in bothbasins. This observation would probably hold for other basins inthe region but cannot be generalized to a larger scale.
The extrapolation technique based on the weekly linear regres-
sion performing well on an annual basis, let us examine the pro-
duction in the Gatineau System on a weekly time step. Fig. 7 shows
the weekly means while Fig. 8 displays the weekly standard devia-
tions of the energy generated with the horizontal and the vertical
approaches. Recall that the vertical approach relies on 21 hydro-
logic models, meaning that both the average weekly values and
standard deviations are given by an interval. If the production pat-
terns are similar, we can see that the extrapolation of Du Lievre
River discharges tend to overestimate the average weekly produc-
tion during the summer season: the average weekly productions are
close to the upper bound of the structural uncertainty. During the
rest of the year, the opposite behavior is observed. Since the mean
values are most of the time within the structural uncertainty range,
we can conclude that the extent of the weekly productions derived
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 120
10
0
10
20Area Ratio
Rel
ativ
e di
ffere
nce
(%)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 120
10
0
10
20Linear Regression
Rel
ativ
e di
ffere
nce
(%)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 120
10
0
10
20Change in mean
Rel
ativ
e di
ffere
nce
(%)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 120
10
0
10
20Autoregressive Model
Rel
ativ
e di
ffere
nce
(%)
Non exceedance probability [ ]
Fig. 6. Annual energy generation—relative differences between the distribution functions
Table 3. Root-Mean-Square Deviations (TWh) between the FourExtrapolation Techniques and the Benchmark
Probabilityrange (%)
Arearatio
Changein mean
Linearregression AR model
100 0.81 1.09 0.61 0.7190 0.65 1.01 0.31 0.6780 0.59 0.85 0.22 0.5850 0.62 1.18 0.29 0.46
© ASCE 04016081-7 J. Water Resour. Plann. Manage.
J. Water Resour. Plann. Manage., 04016081
Dow
nloa
ded
from
asc
elib
rary
.org
by
UN
IVE
RSI
TE
LA
VA
L o
n 11
/14/
16. C
opyr
ight
ASC
E. F
or p
erso
nal u
se o
nly;
all
righ
ts r
eser
ved.
67
with a horizontal approach are consistent with that obtained with a
diversity of hydrologic models.The erratic rainfalls that characterize the beginning of the winter
season result in an increased variability of the production as indi-
cated on Fig. 8. This also reflects in the structural uncertainty,
which is larger during the second half of the year. As we can also
see from the examination of Fig. 8, the standard deviation associ-
ated with the horizontal approach lies most of the time within the
range defined by the structural uncertainty. It means that the
horizontal approach yields weekly energy productions whose
variability is consistent with that obtained with a diversity of
hydrologic models.Fig. 9 shows the average drawdown-refill cycle of the
Baskatong Reservoir when the inflows are derived from the linear
regression. As one can see, the drawdown starts in the summer and
goes on until weeks 12–13 when the snow melt season starts and
the reservoir is quickly refilled at the beginning of the next summer.
For illustrative purposes, the drawdown refill cycle corresponding
to current hydrometeorological conditions are also displayed on thesame figure. We can see that for the selected emission scenario, therefill phase starts two weeks earlier following earlier snowmelt.Under that scenario, weekly river discharges during the first halfof the year (winter) are larger than under the current climate, mean-ing that the reservoir must be depleted at little bit further before thespring freshet. In the summer and the fall, lower rainfalls leadto a steady decline in the storage levels. The reapportion of thisreservoir is an example of nonstructural adaptation measure tomitigate the impacts of a changing flow regime (for a givenemission scenario).
Conclusion
This paper examines the issue of assessing the impact of climatechange in a river basin (target river basin) for which hydrologicprojections have not yet been calculated but are available in neigh-bouring river basins (source river basins). A methodology ispresented to determine whether or not the extrapolation of the pro-jections already available can be used as a substitute to the verticalapproach to climate change assessment. The methodology com-pares the performance of a hydropower system with the streamflowseries derived either directly from a vertical approach or indirectlyby extrapolating the projections available outside the system of in-terest. To account for the structural hydrologic uncertainty inherentto the system, a multimodel approach is implemented within thevertical chain of hydro-climate models. To account of the naturaluncertainty, a stochastic water resource model is developed todetermine the optimal allocation policies using either the vertically-obtained or the extrapolated river discharges. Different extrapola-tion techniques are tested and the performance of the systemcompared to the benchmark, i.e., the statistical distribution ofenergy generation obtained with the vertical chain of models. Ahydropower system in Quebec, Canada, is used as a case studyand the analysis is done for a particular emission scenario.
The analysis of simulation results reveals that the statistical dis-tribution of energy generation can already be well approachedusing an extrapolation technique based on linear regressions.Assuming that current linear relationships between weekly riverdischarges in both basins will not be affected by climate change,simple linear regressions can be used to derive future discharges inthe target river basin from the projected discharges in the sourceriver basin. Annual and weekly productions are within their respec-tive structural uncertainty ranges, indicating that the extrapolationprovides a reasonable approximation of the amount of energy thatcan be expected from the system. This is probably particular to thecase study and cannot be generalized to the rest of Quebec (andbeyond). However, similar tests should be carried out to identifywhat extrapolation technique (assumption) is best for each hydro-logic region in Quebec. Then, the results of the vertical approach
10 20 30 40 500
20
40
60
80
Week
Ene
rgy
(GW
h)
Structural uncertainty
Horizontal
Fig. 7. Energy generation—weekly averages
5 10 15 20 25 30 35 40 45 500
5
10
15
20
25
Week
Ene
rgy
(GW
h)
Structural uncertaintyHorizontal
Fig. 8. Energy generation—weekly standard deviations
4 8 12 16 20 24 28 32 36 40 44 48 52
500
1000
1500
2000
2500
3000
Sto
rage
[hm
3 ]
Expected storage value
Weeks
Current Condition
Climate change Horizontal app
Min Storage
Max Storge
Fig. 9. Baskatong drawdown-refill cycle—horizontal approach versus current conditions
© ASCE 04016081-8 J. Water Resour. Plann. Manage.
J. Water Resour. Plann. Manage., 04016081
Dow
nloa
ded
from
asc
elib
rary
.org
by
UN
IVE
RSI
TE
LA
VA
L o
n 11
/14/
16. C
opyr
ight
ASC
E. F
or p
erso
nal u
se o
nly;
all
righ
ts r
eser
ved.
68
implemented on a particular system could be extrapolated to therest of the hydrologic region encompassing that system.
Appendix. DDP and MPAR(p)
The built-in hydrologic model found is SDDP is a multisite, multi-period autoregressive model with cross-correlated residuals
�
qtðjÞ − μtðjÞσtðjÞ
�
¼X
Pm
i¼1
ϕi;tðjÞ�
qt−iðjÞ − μt−iðjÞσt−iðjÞ
�
þ ϵtðjÞ ð17Þ
where j = site’s index (j ¼ 1; 2; : : : ; J); qt = river discharge at timet (t ¼ 1,2; : : : ; 52 weeks); μt and σt = periodic mean and standarddeviation; ϕi;tðjÞ = lag-i autoregressive parameter at site j and timet; and ξt = time independent stochastic noise with mean zero andperiodic variance σ2
ϵ;tðjÞ. Eq. (17) is compatible with the Benders’decomposition scheme on which SDDP relies to mitigate the curseof dimensionality associate with DP formulations.
The set of parameters to be determined in the model are μtðjÞ,σtðjÞ, ϕi;tðjÞ, and σ2
ϵ;tðjÞ. Given N years of historical (reference) orprojected (future) incremental inflow data qtðjÞ available at site j,the parameters μtðjÞ, σtðjÞ can be estimated by the first and secondmoment as
μqðjÞ ¼1
N
X
N
v¼1
qv;tðjÞσ2t ðjÞ ¼
1
N − 1
X
N
v¼1
½qv;tðjÞ − μtðjÞ�2 ð18Þ
The temporal Lag-k autocorrelation ρtðkÞ between qt andqt−k for each season t is obtained by multiplying Eq. (17) byf½qt−kðjÞ − μt−kðjÞ�=σt−kðjÞg and taking the expectation termby term
ρtðkÞ ¼ ϕ1;tðjÞρt−1jk − 1j þ ϕ2;tðjÞρt−2jk − 2jþ ϕPm;t
ðjÞρt−Pmjk−Pmj ð19Þ
Eq. (19) constitutes a set of Pm linear equations, fork ¼ 1; 2; : : : ;Pm
ρtð1Þ¼ϕ1;tρt−1ð0Þþϕ2;tρt−2ð1Þþ : : :þϕPm;tρt−Pm
ð1−PmÞρtð2Þϕ1;tρt−1ð1Þþϕ2;tρt−2ð0Þþ : : :þϕPm;t
ρt−Pmð2−PmÞ
..
.
ρtðpmÞ¼ϕt;1ρt−1ðPm−1Þþϕt;2ρt−2ðPm−2Þþ : : :þϕt;Pmρt−Pm
ð0Þð20Þ
which may be written as
Pq;tðjÞ · ϕi;tðjÞ ¼ ρi;tðjÞ ð21Þ
where Pq;tðjÞ = periodic covariance matrix given as
2
6
6
6
6
4
1 ρt−1ð1Þ : : : ρt−1ðPm − 1Þρt−1ð1Þ 1 : : : ρt−2ðPm − 2Þ
..
. ... ..
.
ρt−1ðPm − 1Þ ρt−2ðPm − 2Þ : : : 1
3
7
7
7
7
5
and ϕi;tðjÞ and ρi;tðjÞ, respectively, are the vectors of autoregressivecoefficients of the MPAR(P) and autocorrelation coefficients.
ϕi;tðjÞ ¼
2
6
6
6
6
6
6
6
6
4
ϕ1;t
ϕ2;t
ϕ3;t
..
.
ϕPm;t
3
7
7
7
7
7
7
7
7
5
ρi;tðjÞ ¼
2
6
6
6
6
6
6
6
6
4
ρtð1Þρtð2Þρtð3Þ...
ρtðPmÞ
3
7
7
7
7
7
7
7
7
5
According to Salas et al. (1980) the residual variance for eacht, σ2
ϵ;tðjÞ can be estimated as a function of the periodic autore-gressive coefficients ϕi;tðjÞ and the periodic autocorrelationcoefficients ρi;tðjÞ
σ2ϵ;tðjÞ ¼ 1 −
X
Pm
j¼1
ϕj;tðjÞρj;tðjÞ ð22Þ
When the hydrological state variable in SDDP is denoted bythe previous p inflows the one-stage optimization problem canbe written as
Ftðst; qt−1; qt−2; :::; qt−PmÞ ¼ maxfαtbtðst; qt; rtÞ þ Ftþ1g ð23Þ
subject to
stþ1 þ rt þ lt ¼ st þ qt ð24Þ
Ftþ1 − φltþ1stþ1 ≤ γltþ1,1qt þ γltþ1,2qt−1 þ : : : þ γltþ1;Pm
qtþ1−Pmþ βl
tþ1
..
.
Ftþ1 − φLtþ1stþ1 ≤ γLtþ1,1qt þ γLtþ1,2qt−1 þ : : : þ γLtþ1;Pm
qtþ1−Pmþ βL
tþ1 ð25Þ
The parameters φtþ1; γtþ1,1; γtþ1,2; : : : ; γtþ1;Pmand βtþ1 of the
linear segments of Ftþ1 can be calculated from the primal and the
dual information available at the optimal solution of the one stage
optimization problem. According to the Kuhn-Tucker conditions
for optimality the derivative of the objective function with respectto the state variables x is given by
∂F
∂xi¼
X
λi
∂gi
∂xið26Þ
© ASCE 04016081-9 J. Water Resour. Plann. Manage.
J. Water Resour. Plann. Manage., 04016081
Dow
nloa
ded
from
asc
elib
rary
.org
by
UN
IVE
RSI
TE
LA
VA
L o
n 11
/14/
16. C
opyr
ight
ASC
E. F
or p
erso
nal u
se o
nly;
all
righ
ts r
eser
ved.
69
where λi = dual information of the optimization problem; and gi =linear constraintsin Eqs. (24) and (25).
∂Ft
∂xi¼ λw;t
∂
∂xiðstþ1 þ rt þ lt − st − qtÞ
þX
L
l¼1
λlc;t
∂
∂xiðFtþ1 − φl
tþ1stþ1 − γltþ1,1qkt
− γltþ1,2qt−1− γltþ1,3qt−2 − : : : − γltþ1;Pmqtþ1−Pm
− βltþ1Þð27Þ
By taking the partial derivative with respect to the state variablesst; qt−1; qt−2; : : : ; qt−Pm
, the change of the original function Ft
respect the state variable associated to the constraints, isdetermined as
∂Ft
∂st¼ λw;t ð28Þ
∂Ft
∂qt−1¼ λw;t
∂qt
∂qt−1þX
L
l¼1
λlc;t
�
γltþ1,1∂qt
∂qt−1þ γltþ1,2
∂qt−1
∂qt−1
�
∂Ft
∂qt−2¼ λw;t
∂qt
∂qt−2þX
L
l¼1
λlc;t
�
γltþ1,1∂qt
∂qt−2þ γltþ1,3
∂qt−2
∂qt−2
�
..
.
∂Ft
∂qt−Pm
¼ λw;t∂qt
∂qt−Pm
þX
L
l¼1
λlc;t
�
γltþ1,1∂qt
∂qt−Pm
�
ð29Þ
Now, let us say that at stage t; s°t; q°t−1; q
°t−2; : : : ; q
°t−Pm
aresampled and, in order to include the stochasticity of the problem,K vector of inflows qKt are generated by using the periodic autor-egressive model Eq. (17).
Having into account that Fkt , which will be added to the ex-
pected cost-to-go function at stage ðt − 1Þ, can be approximated by
Fkt ≤ φl;k
t s°t þ γl;kt;1q°t−1 þ γl;kt;2q
°t−2 þ : : : þ γl;kt;Pm
q°t−Pmþ βl;k
t
ð30Þ
The slopes φl;kt ; γl;kt;1; γ
l;kt;2; :::; γ
l;kt;Pm
are determined using the dualinformation of the optimal solution at stage t Eqs. (28) and (29), foreach k scenario as
∂Fkt
∂st¼ φl;k
t ðjÞ ¼ λkw;tðjÞ ð31Þ
∂FKt
∂qt−1¼ γl;kt;1ðjÞ ¼
�
λl;kw;tðjÞ þX
L
l¼1
λl;kc;tγltþ1,1ðjÞ
�
∂qt
∂qt−1þX
L
l¼1
λl;kc;tγltþ1,2ðjÞ
∂FKt
∂qt−2¼ γl;kt;2ðjÞ ¼
�
λl;kw;tðjÞ þX
L
l¼1
λl;kc;tγltþ1,1ðjÞ
�
∂qt
∂qt−2þX
L
l¼1
λl;kc;tγltþ1,3ðjÞ
..
.
∂FKt
∂qt−Pm
¼ γl;kt;PmðjÞ ¼
�
λl;kw;t þ
X
L
l¼1
λl;kc;tγ
ltþ1,1
�
∂qt
∂qt−Pm
ð32Þ
Defining Γ as
ΓtðjÞ ¼ λl;kw;tðjÞ þ
X
L
l¼1
λl;kc;tγltþ1,1ðjÞ ð33Þ
and by using Eq. (17) to find the derivatives of qt respect the hydro-logic variables, the set of Eq. (31) can be rewritten as
γl;kt;1ðjÞ ¼ ΓtðjÞσtðjÞσt−1ðjÞ
ϕt;1ðjÞ þX
L
l¼1
λl;kc;tγltþ1,2ðjÞ ð34Þ
γl;kt;2ðjÞ ¼ ΓtðjÞσtðjÞσt−2ðÞj
ϕt;2ðjÞ þX
L
l¼1
λl;kc;tþ1γ
ltþ1,3ðjÞ ð35Þ
γl;kt;PmðjÞ ¼ ΓtðjÞ
σtðjÞσt−Pm
ðjÞϕt;PmðjÞ ð36Þ
Taking the expectation over the K artificially generated flows,the slope vectors φl
t; γlt;1; γ
lt;2; : : : ; γ
lt;Pm
can be determined
φltðjÞ ¼
1
K
X
K
k¼1
λkw;tðjÞ ð37Þ
γlt;pðjÞ ¼1
K
X
K
k¼1
γl;kt;pðjÞ; p ¼ 1,2, : : : ,Pm ð38Þ
The jth element of the vector constant term is given by
βltðjÞ ¼
1
K
X
K
k¼1
Fkt − φl
tðjÞs°tðjÞ − γt;1ðjÞlq°t−1ðjÞ
− γlt;2ðjÞq°t−2ðjÞ− · · · −γlt;PmðjÞq°t−Pm
ðjÞ ð39Þ
© ASCE 04016081-10 J. Water Resour. Plann. Manage.
J. Water Resour. Plann. Manage., 04016081
Dow
nloa
ded
from
asc
elib
rary
.org
by
UN
IVE
RSI
TE
LA
VA
L o
n 11
/14/
16. C
opyr
ight
ASC
E. F
or p
erso
nal u
se o
nly;
all
righ
ts r
eser
ved.
70
Acknowledgments
The authors acknowledge NSERC, Ouranos, Hydro-Quebec, andMDDELCC. We also thank the reviewers for their constructivecomments and suggestions. All SDDP-derived simulation resultscan be accessed through the U. Laval webpage of the second author(https://www.gci.ulaval.ca/departement-et-professeurs/professeurs-et-personnel/professeurs/fiche/show/tilmant-amaury/models-and-data-518/). A READ-ME.txt file describes the Matlab structurecontaining the results.
References
Botzen, W., and van der Bergh, J. (2008). “Insurance against climatechange and flooding in the Netherlands: Present, future, and compari-son with other countries.” Risk Anal., 28(2), 413–426.
Brown, C., Ghile, Y., Laverty, M., and Li, K. (2012). “Decision scaling:Linking bottom-up vulnerability analysis with climate projections inthe water sector.” Water Resour. Res., 48(9), W09537.
Brown, C., Werick, W., Leger, W., and Fay, D. (2011). “A decision analyticapproach to managing climate risks: Application to the upper greatlakes.” J. Am. Water Resour. Assoc., 47(3), 524–534.
Carpenter, T. M., and Georgakakos, K. P. (2001). “Assessment of FolsomLake response to historical and potential future climate scenarios. 1:Forecasting.” J. Hydrol., 249(1–4), 148–175.
CEHQ (Centre d’Expertise Hydrique du Québec). (2012). “Délimitationdes bassins versants correspondant aux stations hydrométriquesouvertes et fermées.” ⟨ftp://ftp.mddep.gouv.qc.ca/⟩ (Jun. 18, 2015).
Chen, J., Brissette, F. P., Poulin, A., and Leconte, R. (2011). “Overalluncertainty study of the hydrological impacts of climate change fora Canadian Watershed.” Water Resour. Res., 47(12), W12509.
Dessai, S., and Hulme, M. (2004). “Does climate adaptation policy needprobabilities?” Clim. Policy, 4(2), 107–128.
Duan, Q., Gupta, V., and Sorooshian, S. (1993). “Shuffled complexevolution approach for effective and efficient global minimization.”J. Optim. Theory Appl., 76(3), 501–521.
Fortin, J., Turcotte, R., and Massicotte, R. (2001). “Distributed watershedmodel compatible with remote sensing and GIS data. I: Description ofmodel.” J. Hydrol. Eng., 10.1061/(ASCE)1084-0699(2001)6:2(91),91–99.
Goor, Q., Kelman, R., and Tilmant, A. (2011). “Optimal multipurpose mul-tireservoir operation model with variable productivity of hydropowerplants.” J. Water Resour. Plann. Manage., 10.1061/(ASCE)WR.1943-5452.0000117, 258–267.
Harrison, G. P., and Whittington, H. W. (2002). “Susceptibility of theBatoka Gorge hydroelectric scheme to climate change.” J. Hydrol.,264(1–4), 230–241.
Iglesias, A., and Garrote, L. (2015). “Adaptation strategies for agriculturalwater management under climate change in Europe.” Agric. Water
Manage., 155, 113–124.IPCC (Intergovernmental Panel on Climate Change). (2013). “Summary
for policymakers. Climate change 2013: The physical science basis.”Contribution of Working Group I to the 5th Assessment Rep. of the
Intergovernmental Panel on Climate Change, T. F. Stocker, et al.,eds., Cambridge University Press, Cambridge, U.K.
Jeuland, M., and Whittington, D. (2014). “Water resources planning underclimate change: Assessing the robustness of real options for the blueNile.” Water Resour. Res., 50(3), 2086–2107.
Klemeš, V. (1986). “Operational testing of hydrological simulation
models.” Hydrol. Sci. J., 31(1), 13–24.
Lempert, R., and Schlesinger, M. (2000). “Robust strategies for abating
climate change.” Clim. Change, 45(3–4), 387–401.
Music, B., and Caya, D. (2007). “Evaluation of the hydrological cycle over
the Mississippi River Basin as simulated by the Canadian regional
climate model (CRCM).” J. Hydrometeorol., 8(5), 969–988.
Nash, J., and Sutcliffe, J. (1970). “River flow forecasting through
conceptual models. Part I: A discussion of principles.” J. Hydrol.,
10(3), 282–290.
New, M., Lopez, A., Dessai, S., and Wilby, R. (2007). “Challenges in using
probabilistic climate change information for impact assessments: An
example from the water sector.” Philos. Trans. R. Soc. London A,
365(1857), 2117–2131.
Oudin, L., Hervieu, F., Michel, C., Perrin, C., Andréassian, V., Anctil, F.,
and Loumagne, C. (2005). “Which potential evapotranspiration input
for a lumped rainfallrunoff model? Part 2—Towards a simple and effi-
cient potential evapotranspiration model for rainfall-runoff modelling.”
J. Hydrol., 303(1–4), 290–306.
Perrin, C. (2000). “Vers une amélioration dun modèle global pluie-debit.”
Ph.D. thesis, Institut National Polytechnique de Grenoble-INPG,
Grenoble, France.
Prudhomme, C., Wilby, R., Crooks, S., Kay, A., and Reynard, N. (2010).
“Scenario-neutral approach to climate change impact studies: Applica-
tion to flood risk.” J. Hydrol., 390(3–4), 198–209.
Salas, J., Delleur, J., Yevjevich, V., and Lane, W. (1980). Applied modeling
of hydrologic time series, Water Resources Publications, Littleton, CO.
Seiller, G., Anctil, F., and Perrin, C. (2012). “Multimodel evaluation
of twenty lumped hydrological models under contrasted climate
conditions.” Hydrol. Earth Syst. Sci., 16(4), 1171–1189.
Thiboult, A., and Anctil, F. (2015). “Assessment of a multimodel ensemble
against an operational hydrological forecasting system.” Can. Water
Resour. J., 40(3), 272–284.
Tilmant, A., Pinte, D., and Goor, Q. (2008). “Assessing marginal water
values in multipurpose multireservoir systems via stochastic
programming.” Water Resour. Res., 44(12), W12431.
Turcotte, R., Lacombe, P., Dimnik, C., and Villeneuve, J.-P. (2004).
“Prévision hydrologique distribuée pour la gestion des barrages publics
du Québec.” Can. J. Civ. Eng., 31(2), 308–320.
Valéry, A., Andréassian, V., and Perrin, C. (2014). “Catchments, ‘As simple
as possible but not simpler’: What is useful in a temperature-based
snow-accounting routine? Part 2—Sensitivity analysis of the Cemaneige
snow accounting routine on 380.” J. Hydrol., 517, 1176–1187.
Vicuna, S., Dracup, J. A., Lund, J. R., Dale, L. L., and Maurer, E. P. (2009).
“Basin scale water system operations with uncertain future climate
conditions: Methodology and case studies.” Water Resour. Res.,
46(4), 1–19.
Weaver, C. P., Lempert, R. J., Brown, C., Hall, J. A., Revell, D., and
Sarewitz, D. (2013). “Improving the contribution of climate model
information to decision making: The value and demands of robust
decision frameworks.” Wiley Interdiscip. Rev. Clim. Change, 4(1),
39–60.
Wilby, R., and Dessai, S. (2010). “Robust adaptation to climate change.”
Weather, 65(7), 180–185.
Yao, H., and Georgakakos, A. (2001). “Assessment of Folsom Lake
response to historical and potential future climate scenarios. 2:
Reservoir management.” J. Hydrol., 249(1–4), 176–196.
© ASCE 04016081-11 J. Water Resour. Plann. Manage.
J. Water Resour. Plann. Manage., 04016081
Dow
nloa
ded
from
asc
elib
rary
.org
by
UN
IVE
RSI
TE
LA
VA
L o
n 11
/14/
16. C
opyr
ight
ASC
E. F
or p
erso
nal u
se o
nly;
all
righ
ts r
eser
ved.
71
Appendix B
Paper II:Optimizing multireservoirsystem operating policies using exogenoushydrologic variables
Résumé
La programmation dynamique stochastique duale (SDDP) fait partie des quelques algorithmes disponibles
pour optimiser les politiques de gestion des systèmes hydroélectriques de grande taille. Ce manuscrit
présente une nouvelle extension, appelée SDDPX, dans laquelle des variables hydrologiques exogènes
telles que l’équivalent de neige en eau et/ou la température de la surface de la mer, sont incluses dans
le vecteur d’état avec les variables traditionnelles (endogènes), comme les débits des pas de temps
précédents. Avec cette extension, les décisions de relâche d’eau sont maintenant fonction des stocks,
des débits des pas de temps précédents et des variables exogènes pertinentes à même de représenter
des processus hydrologiques plus complexes que ceux des formulations SDDP traditionnelles. Pour
illustrer le gain potentiel associé à l’utilisation de variables exogènes pour la gestion d’un système à
réservoir multiples, le système hydroélectrique de Rio Tinto (RT) est utilisé comme cas d’étude. Il est
situé dans le bassin du fleuve Saguenay-Lac-St-Jean au Québec (Canada), et dispose d’une capacité
installée de 3137 MW. La performance du système est évaluée pour différentes combinaisons de vari-
ables d’état hydrologiques, allant du modèle autorégressif d’ordre 1 à une formulation plus complexe
impliquant les débits des pas de temps précédents, l’équivalent de neige en eau et les précipitations
hivernales. Une procédure de ré-optimisation est implémentée dans laquelle les fonctions de béné-
fices futurs dérivées de SDDPX sont utilisées au sein d’une simulation réalisée sur l’historique des
variables hydrologiques endogènes et exogènes.
http://onlinelibrary.wiley.com/doi/10.1002/2017WR021701/abstract.
72
RESEARCH ARTICLE10.1002/2017WR021701
Optimizing Multireservoir System Operating Policies UsingExogenous Hydrologic VariablesJasson Pina1, Amaury Tilmant1 , and Pascal Cot�e2
1Department of Civil Engineering and Water Engineering, Universit�e Laval, QC, Canada, 2Rio Tinto, Quebec PowerOperation, Jonqui�ere, QC, Canada
Abstract Stochastic dual dynamic programming (SDDP) is one of the few available algorithms tooptimize the operating policies of large-scale hydropower systems. This paper presents a variant, calledSDDPX, in which exogenous hydrologic variables, such as snow water equivalent and/or sea surfacetemperature, are included in the state space vector together with the traditional (endogenous) variables,i.e., past inflows. A reoptimization procedure is also proposed in which SDDPX-derived benefit-to-gofunctions are employed within a simulation carried out over the historical record of both the endogenousand exogenous hydrologic variables. In SDDPX, release policies are now a function of storages, past inflows,and relevant exogenous variables that potentially capture more complex hydrological processes than thosefound in traditional SDDP formulations. To illustrate the potential gain associated with the use of exogenousvariables when operating a multireservoir system, the 3,137 MW hydropower system of Rio Tinto (RT)located in the Saguenay-Lac-St-Jean River Basin in Quebec (Canada) is used as a case study. Theperformance of the system is assessed for various combinations of hydrologic state variables, ranging fromthe simple lag-one autoregressive model to more complex formulations involving past inflows, snow waterequivalent, and winter precipitation.
1. Introduction
The operation of a multireservoir system is a complex, multistage, stochastic decision-making probleminvolving, among others, (i) a trade-off between immediate and future consequences of a release decision,(ii) considerable risks and uncertainties, and (iii) multiple objectives and operational constraints (Oliveira &Loucks, 1997). For hydropower systems, the problem can also be nonlinear because the production ofhydroelectricity depends on the product of the outflow and the head (storage). This decision-making prob-lem has been studied for several decades and state-of-the-art reviews can be found in Yeh (1985), Labadie(2004), Rani and Moreira (2009)], and more recently in Ahmad et al. (2014).
Dynamic programming (DP), first introduced by Bellman (1957), has been one of the most popular optimiza-tion techniques to determine reservoir operating policies. The method solves the problem by decomposingthe multistage decision-making problem into simpler one-stage problems, which are then solved recur-sively. DP can be easily expanded to accommodate the stochasticity of the hydrologic input by adding ahydrologic variable to the state space vector. The resulting stochastic DP (SDP) formulation, often referredto as a Markov decision process, explicitly considers the streamflow lag-1 correlation found in the flowrecords. SDP solves the problem by discretizing stochastic variables, as well as the system status, to obtainan optimal solution for each discrete value of the state space that characterizes the system (Tejada-Guibertet al., 1995). Although conceptually attractive, SDP is however limited by the so-called curse of dimensional-ity, which limits its application to small systems involving no more than four state variables (storage andhydrologic).
Incorporating more hydrologic information in the state vector has the potential to enhance SDP-derivedpolicies and thus improve the performance of the system. Various approaches have been proposed the liter-ature to address this challenge. For example, Bras et al. (1983) combine real time forecasts with an adaptivecontrol technique in SDP to update flow transition probabilities. Stedinger et al. (1984) develop an SDPmodel which employs the best inflow forecast for the current period’s flow to define the policy. Kelmanet al. (1990) propose sampling SDP (SSDP) to better capture the complex temporal and spatial structures of
Key Points:� The SDDP algorithm can handle
multiple endogenous and exogenoushydrologic state variables� Gains in energy production can be
observed when more hydrologicvariables are included in the statespace vector� The marginal water values tend to
increase while spillage losses arereduced regardless of the hydrologicstatus of the system
Correspondence to:A. Tilmant,[email protected]
Citation:Pina, J., Tilmant, A., & Cot�e, P. (2017).Optimizing multireservoir systemoperating policies using exogenoushydrologic variables. Water ResourcesResearch, 53. https://doi.org/10.1002/2017WR021701
Received 19 AUG 2017
Accepted 10 NOV 2017
Accepted article online 15 NOV 2017
VC 2017. American Geophysical Union.
All Rights Reserved.
PINA ET AL. SDDPX EXOGENOUS VARIABLES 1
Water Resources Research
PUBLICATIONS
73
the streamflow process. Karamouz and Vasiliadis (1992) propose another alternative to SDP, coinedBayesian SDP (BSDP), in which prior flow transition probabilities are regularly updated using the Bayes’theorem. Tejada-Guibert et al. (1995) determine the value of hydrological information for several SDPformulations each employing a different set of hydrologic state variables and, similarly, Kim and Palmer(1997) illustrate the potential advantage of using the seasonal flow forecasts and other hydrologic informa-tion by comparing the performance of the BSDP against alternative SDP formulations. Furthermore, Faberand Stedinger (2001) and Kim et al. (2007) examine the use Ensemble Streamflow Prediction (ESP) withinthe SSDP optimization framework. Cot�e et al. (2011) incorporate a new hydrologic state variable in SSDP asa linear combination of snow water equivalent and soil moisture, and more recently, Desreumaux et al.(2014) present the effect of using various hydrological variables on SDP-derived policies of the Kemanohydropower system in British Columbia. Another active research line focuses on the incorporation of climatevariables such as the El Nino Southern Oscillation (ENSO) and Pacific Decadal Oscillation (PDO) climate sig-nals into reservoir operation models (Gelati et al., 2013; Hamlet & Lettenmaier, 1999; Kwon et al., 2009).However, most of the above-mentioned studies are limited to small-scale problems, meaning that a trade-off must be found between the complexity of system to be studied (number of reservoirs) and the complex-ity of the hydrologic processes that can be captured.
This trade-off between system and hydrologic complexities can largely be removed using stochastic dualDP (SDDP), an extension of SDP that is not affected by the curse of dimensionality (Pereira & Pinto, 1991).To achieve this, SDDP builds a locally accurate approximation of the expected-benefit-to-go functionsthrough piecewise linear functions. With such an approximation, there is no need to evaluate the functionover a dense grid; the benefits can now be derived from extrapolation and not interpolation as in SDP. Aswe will see later, these piecewise linear functions are constructed from the primal and dual solutions of theone-stage optimization problem and can be interpreted as Benders cuts in a stochastic, multistage decom-position algorithm. To increase the accuracy of the approximation, SDDP uses an iterative procedure inwhich new cuts are added to the most interesting region of the state space until the approximation is statis-tically acceptable. As explained in Tilmant and Kelman (2007), to implement the decomposition scheme,the one-stage optimization problem must be formulated as a convex problem, such as a linear program.SDDP has largely been used in hydropower-dominated systems such as Norway (Gjelsvik et al., 2010; Moet al., 2001; Rotting & Gjelsvik, 1992), South and Central America (Homen-de Mello et al., 2011; Pereira, 1989;Shapiro et al., 2013), and New Zealand (Kristiansen, 2004). The SDDP algorithm constitutes the coreof generic hydro-economic models that have been used to analyze a variety of policy issues in theEuphrates-Tigris River basin (Tilmant et al., 2008), the Nile River basin (Goor et al., 2011), the Zambezi Riverbasin (Tilmant & Kinzelbach, 2012), or in Spain (Macian-Sorribes et al., 2016; Pereira-Cardenal et al., 2016).
In SDDP, the hydrologic uncertainty is typically captured through a multisite periodic autoregressive modelMPAR(p). This model is capable of representing seasonality, serial, and spatial streamflow correlations withina river basin and among different basins. It is also needed to analytically derive some of the parameters ofthe linear segments used to approximate the benefit-to-go functions, and to produce synthetic streamflowsscenarios for the simulation phase of this iterative algorithm. Furthermore, the convexity requirement ofSDDP is guaranteed because the MPAR(p) is linear. Recent works reveal an interest in improving the built-inhydrological model. In Lohmann et al. (2015), a new approach to include spatial information is presented.Pritchard (2015) models inflows as a continuous process with a discrete random innovation. Poorsepahy-Samian et al. (2016) propose a methodology to estimate the cuts parameters when a Box-Cox transforma-tion is used to normalize inflows, and more recently, Raso et al. (2017) present a streamflow model with amultiplicative stochastic component and a nonuniform time step.
This study focuses on the incorporation of new hydrologic information into the SDDP algorithm. This newinformation, encapsulated as additional, exogenous, hydrologic state variables through a built-in MPARXmodel, aims at better capturing the hydrologic processes responsible for reservoir inflows. By definition,reservoir inflows (the endogenous variables) are causal dependent on the exogenous hydrologic state varia-bles while the opposite is not true. For example, winter snow pack, precipitation, or sea surface temperatureare exogenous to river discharges. Incorporating exogenous hydrologic state variables into SDDP requiresthat some aspects of the iterative algorithm be modified. Particular attention is devoted to the new mathe-matical formulation of the cuts approximating the benefit-to-go functions, and how the correspondingparameters can be analytically derived from the primal and dual information that become available as the
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 2
74
algorithm progresses backward (backward optimization phase). In the forward simulation phase, due to thestructure of the built-in hydrologic model (MPARX), the number of hydrologic sequences is now limited tothe length of hydrological records/direct measurements.
Following the temporal decomposition approach Zahraie and Karamouz (2004), the proposed SDDPX for-mulation solves the midterm hydropower scheduling problem and therefore seek to properly capture themidterm to long-term hydrologic uncertainties. Since the nonlinear nonconvex hydropower function can-not be used as such in the SDDP algorithm, usually simplifications of the function are used. We can rely, forexample, on a production coefficient (Archibald et al., 1999), convex hull approximations (Goor et al., 2011),McCormick envelopes (Cerisola et al., 2012), or a concave approximation (Zhao et al., 2014). To deal withthe nonconvexity of the optimization problem, the hydropower production functions are approximated byconvex hulls. This new formulation is illustrated with the 3,137 MW hydropower system of Rio Tinto (RT)located in the Saguenay-Lac-St-Jean River Basin in Quebec (Canada). The five hydropower plants have thecapacity to produce approximately 90% of the electricity required for the production of aluminum, forcingRT to purchase energy to fully ensure the production of the mineral. Therefore, a joint optimization of physi-cal (power plants) and financial (contracts) assets is developed using SDDPX, and a comparative analysis ofthe performance of the system with various combinations of endogenous and exogenous hydrologic statevariables is performed.
The paper is organized as follows: section 2 starts with a presentation of the reservoir operation problem,which is then followed by a description of the SDDP algorithm and its variant SDDPX with exogenoushydrologic variables. This section ends with a presentation of the case study. Afterward, optimization resultsare discussed in section 3. Finally, concluding remarks are given in section 4.
2. Materials and Methods
2.1. The Hydroelectric Reservoir Operation ProblemThe operation of a multireservoir system is a multistage decision-making problem. When framed as an opti-mization problem, the goal is to determine a sequence of optimal decisions xt that maximizes the expectedsum of net benefits from system operation Z over a given planning period. Let T be the number of stages inthe planning period, btð�Þ be the net benefit function at stage t, mð�Þ be the terminal value function, E½�� bethe expectation operator, and St be a vector of state variables characterizing the system at the beginning ofstage t, the objective function can be written as:
Z5EXT
t51
btðSt; xtÞ1mðST11Þ" #
(1)
This objective function will be maximized to the extent made possible by operational and/or institutionalconstraints affecting the state and decision variables.
2.2. One-Stage SDDP ProblemSDDP solves the optimization problem (1) by decomposing it into a sequence of one-stage problems thatare solved recursively. Let us first adopt the same notation as in Tilmant et al. (2008) and say that the waterresources system is represented by a network with J nodes (e.g., reservoir, power plant). Imagine that theobjective is to maximize the net benefits associated to the production of hydroelectricity. The immediatebenefit function btð�Þ includes the net benefits from hydropower generation and penalties for not meetingtarget demands and/or violating constraints is expressed as
btðSt; xtÞ5XJ
j51
Psys;tðjÞstðptðjÞ2htðjÞÞ� �
�2n0t zt (2)
where st is the number of hours in period t, Psys;t is the vector of power generated (MW), p is the vector ofenergy price ($/MWh), h is the vector of the operation and maintenance cost ($/MWh), and zt is the vectorof deficits or surpluses (e.g., energy demand, environmental flows) penalized by the vector n0t of penalties($/unit). Sales and purchases are included in vector zt and the penalty is the price of energy specified in thecorresponding contract ($/MWh).
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 3
75
At stage t, the system status St is described by the vector of storage st , the hydrological state variable ht ,and the amount left of energy wt in C contracts. The one-stage SDDP optimization problem is expressedas:
Ftðst; ht;wtÞ5maxfbtðst; ht;wt; xtÞ1Ft11g (3)
In SDDP, the benefit-to-go function is represented by the scalar Ft11 which is bounded by a locally accuratelinear approximation. The stage to stage transformation function corresponds to the mass balance equationinvolving the vector of spillage losses lt , the vector of inflows qt , the vector of the turbined outflows rt , thevector of evaporation losses et , and the connectivity matrix CMR
st112CMRðrt1ltÞ5st1qt2etðstÞ (4)
CMRðj; kÞ5 1(21) when reservoir j receives(releases) water from(to) reservoir k.
As mentioned earlier, the benefit-to-go function Ft11, which is represented by a scalar in equation (3), isbounded from above by inequality constraints:
Ft112ult11st112vl
t11wt11 � Clt11ht111bl
t11
�
Ft112uLt11st112vL
t11wt11 � CLt11ht111bL
t11
8>><>>: (5)
where L is the number of cuts. The parameters ut11; vt11; bt11, and Ct11 must have been calculatedfrom the primal and the dual information available at the optimal solution of the one-stage optimiza-tion problem at the stage t 1 1 (Tilmant et al., 2008). In previous SDDP application to water resourcessystems, the hydrological variables are the natural inflows observed during the last p periods htðjÞ5½qt21ðjÞ; qt22ðjÞ; . . . ; qt2pðjÞ� and the current inflow is described by a multisite periodic autoregressivemodel MPAR(p):
qtðjÞ2lqtðjÞ
rqt ðjÞ5Xp
i51
/i;tðjÞqt2iðjÞ2lqt2i
ðjÞrqt2i ðjÞ
� �1�tðjÞ (6)
where lqtand rqt are, respectively, the vectors of periodic mean and standard deviation of qt at period t,
/i;t is the vector of autoregressive coefficients, and �t is the time dependent and spatially correlated sto-chastic noise of zero mean and variance r2
�;t .
The nonlinear hydropower production Psys;t (MW), defined as the product of the net head ht (m), the releasert (m3=s), the turbines/generators efficiency g, and the specific weight of water cw (MN/m3):
Psys;t5cw � gðst; st11; rtÞ � hðst; st11Þ � rt (7)
is approximated by convex hulls and stored in the constraints set (8):
Psys;t2w1st11=22x1rt � w1st=21d1
�
Psys;t2wHst11=22xHrt � wHst=21dH
8>><>>: (8)
where Psys;t is the approximated power, H is the number of linear segments, w, x, and d are the vectors ofparameters determined according to the procedure described in Goor et al. (2011).
The load Dt (MWh) must be met with the energy produced by the system Psys;t and the sales/purchasesthrough the contracts: X
J
Psys;tðjÞst1X
C
utðcÞ1Pp;tDt2Ps;tDt5Dt (9)
Energy transactions can be handled by (i) instant power contracts in which a given amount of power Pp;t
can be bought and surpluses be sold Ps;t during time period Dt, and (ii) purchase energy contracts wtðcÞ. Inthe latter, the amount of energy utðcÞ5Pw;tst that can be withdrawn from the contracts follows the energybalance equation:
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 4
76
wt11ðcÞ1utðcÞ5wtðcÞ (10)
and it is limited by the minimum and maximum instant power withdrawn Pw;t
Pminw st � ut � Pmax
w st (11)
2.3. Incorporating Exogenous Hydrologic Variables into SDDPIncorporating exogenous hydrologic variables into the state space vector of SDDP offers the potential toimprove the performance of SDDP-derived release policies. Using p previous inflows qt and b past exoge-nous variables Xt, the hydrological state variable ht becomes ht5½qt21; qt22,. . .,qt2p; Xt2j,. . ., Xt2b�. Withthese variables, the incremental flow at node j, qtðjÞ, can be derived from a multisite periodic autoregressivemodel with exogenous variables MPARX(p,b) (Ljung, 1999):
qtðjÞ2lqtðjÞ
rqt ðjÞ5Xp
i51
/i;tðjÞqt2iðjÞ2lqt2i
ðjÞrqt2i ðjÞ
� �1Xb
j5i
#j;tðjÞXt2jðjÞ2lXt2j
ðjÞrXt2jðjÞ
� �1�tðjÞ (12)
where lXtand rXt are, respectively, the vectors of the periodic mean and the standard deviation of the
exogenous variables and #j;t is the vector of the exogenous regressors. As indicated in equation (12), theexogenous variables may cover a different range of past input values, from i to b, not necessarily startingfrom t 2 1.
The main modification to the traditional SDDP formulation lies in the calculation of the hyperplanes’ param-eters ut11; vt11; bt11, and Ct11 (see equation (5)). In particular, Ct11 is the vector of linear parameters[ct11;1; ct11;2,. . ., ct11;p; ct11;p1j,. . ., ct11;p1b] associated to ht11 5 ½qt; qt21,. . .,qðt2pÞ11; Xðt2jÞ11,. . ., Xðt2bÞ11�.The procedure to analytically derive the hyperplanes’ parameters when exogenous hydrologic variables areadded to the state space vector is described in Appendix A. This procedure is implemented in the backwardoptimization phase of the SDDP algorithm. The accuracy of the piecewise linear approximation of Ft11 isthen evaluated in a forward simulation phase and if they fail to pass the test, a new backward optimizationphase is implemented (otherwise the algorithm stops). At each iteration, new hyperplanes are added to theconstraints set, refining the approximation of Ft11.
Both phases require different sets of inflows. In the backward phase, K inflows scenarios (backward open-ings) at each node of the system are generated using the MPARX(p,b). Actually, as explained in Appendix A,these scenarios are needed to analytically calculate the hyperplanes’ parameters, and ultimately to derivethe upper bound to the true expected benefit-to-go function. In the forward phase, two different optionsexist to generate the M hydrologic sequences required to simulate the system: (i) one can use theMPARX(p,b) to generate synthetic streamflow sequences based on historical records of both endogenousand exogenous hydrologic variables, (ii) or one can rather rely on series generated outside of SDDPX usingany relevant hydrologic model. Hence, in contrast to the MPAR model used in SDDP, the built-in hydrologicmodel (MPARX) can no longer be used to generate any number of streamflow sequences because the exog-enous variables are independent from river discharges. This might become a limitation if limited hydrologicdata are available.
2.4. Case StudyThe hydroelectric system of Rio Tinto located in the Saguenay-Lac-Saint Jean River Basin, Quebec (Canada)is used as a case study. It includes four reservoirs and five hydropower plants: three on the P�eribonca Riverand two on the Saguenay River. The drainage area is about 78,000 km2. In the northern part of the basin,there are two large reservoirs: Manouane and Passes-Dangereuses with a storage capacity of 2.7 and5.2 km3, respectively. The downstream reservoir system, in which the reservoir Lac-Saint-Jean is included,drains the Basins of the P�eribonka River, Ashuapmushuan River, Mistassini River, and Mistassibi River. Reser-voir Lac-St-jean has important recreational and sport-fishing industries which highly constrain the storagelevels during summer and autumn seasons. Figure 1 shows the reservoir system configuration of the RThydroelectric network and Table 1 lists the main characteristics of the system.
These installations can generate more than 17 TWh/yr, which is roughly 90% of the electricity required forthe production of aluminum, forcing RT to buy energy to fully ensure the production of the mineral. Thereare two contracts (C 5 2): one yearly energy contract and one contract available at the end of the winterseason.
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 5
77
The 53 year hydrologic and climatic series (at nodes 1–5) provided by the Quebec Power Operation Depart-ment of RT are used to parameterize the multisite periodic autoregressive model with exogenous variablesMPARX(p,b) (equation (12)), which residuals �t are fitted to a three-parameter log normal distribution. Forthis comparative analysis, four SDDP formulations, each having a specific hydrologic model, are imple-mented and their performance compared. The first formulation relies on a MPAR(1) model as presented inTilmant et al. (2008). The second formulation attempts at better capturing the temporal persistence of thehydrologic processes through a MPAR(p) model (Matos & Finardi, 2012; Pina et al., 2016). Note that theorder p of the autoregressive model varies in space (node) and time (week). The third and fourth formula-tions are based on the MPARX(p,b) described above, but differ in the selected exogenous variables, whichare, respectively, the snow water equivalent SWEt and the combination of the SWEt and the accumulatedwinter precipitation Pw
t , i.e., winter precipitation from October until time t (Kim & Palmer, 1997). Table 2 liststhe main characteristics of the alternative SDDP formulations.
Figure 1. Rio Tinto hydropower system.
Table 1Rio Tinto Hydropower System Characteristics
Node Id Name Storage (hm3) Capacity (MW)
1 RLM Reservoir Lac-Manouane 2,6572 RPD Reservoir Passes-Dangereuses 5,228 8443 RCD Reservoir Chute du Diable 396 2354 CCS Chute-�a-la-Savane RORa 2505 RLSJ Reservoir Lac-St-Jean 5,083 4546 SH Shipshaw ROR 1,354
aROR: run of the river power plant.
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 6
78
3. Analysis of Simulation Results
All SDDP formulations are implemented on exactly the same hydrocli-matic data: K 5 40 backward openings, M 5 40 hydrological and cli-matic sequences over a planning period of 260 weeks (T 5 260 or5 years). The results are analyzed after reoptimizing the polices alongthe 53 year historical hydrologic sequences. To achieve this, theweekly SDDP-derived benefit-to-go functions of the third year (Ft11;t 5 105, 106,. . ., 157) are used to determine the corresponding releasedecisions by maximizing current plus expected future benefits, sub-ject to system constraints. Retrieving the cuts of the third year is moti-
vated by the fact that the impacts of the boundary conditions (initial storages and zero terminal valuefunctions) on the benefit-to-go functions are minimal when a time period of 2 years is used as a buffer zone(years 1 and 2 for the initial conditions and years 4 and 5 for the terminal conditions). When dealing withsystems with larger carry-over storage capacity, it might be needed to increase the length of the bufferzone by increasing the length of the planning period (e.g., from 5 to 10 years).
To better perceive the advantage of incorporating more hydrologic state variables, the problem is solvedfor two different configurations based on the RT power system. The first system configuration is a simplified,hypothetical, system with a single objective: the maximization of hydropower generation. Important sec-ondary objectives like recreation and flood control as well as the energy load and the contracts are ignoredin order to dedicate all the flexibility offered by the reservoirs to the production of energy. These secondaryobjectives are included in the second configuration which is the actual model of the system. Here the opti-mization is performed on both the physical (power plants, reservoirs) and financial assets (portfolio ofcontracts).
Using SDDP(1) as a benchmark, Table 3 lists, for both configurations, the mean annual reduction in spillagelosses and the annual energy gains one can expect when more hydrologic state variables (endogenous andexogenous) are added to the state space vector. For the second configuration, Table 3 also shows thereduction in energy purchases.
3.1. Maximization of Hydropower GenerationAs we can see in Table 3, increasing the number of lags from 1 to p in the first configuration alreadyincreases the amount of energy produced by 54.41 GWh. Incorporating exogenous variables furtherincreases the production of energy by 46.3 and 50.9 GWh depending on whether SWE or the combinationof SWE and winter precipitation are used. Assuming an average market price of 45 US$/MWh (New YorkIndependent System Operator, 2017), these gains correspond to a 2.45–4.70 million US$ increase in annualenergy value. These energy gains are made possible by improved operating policies that better exploit thestorage capacity of the system.
Figure 2 presents the drawdown-refill cycle of the two largest reservoirs of the system, Passes-DangereusesReservoir (RPD) and Lac-St-Jean Reservoir (RLSJ). The simulated trajectories reveal how the incorporation ofthe exogenous hydrologic variables affects the operating policies during the winter, and the extent towhich the reservoirs must be depleted before the spring snowmelt. In this power system, a large portion ofthe energy comes from run-of-river power stations that are prone to spilling. Consequently, lowering the
Table 2SDDP Formulations
MPARX(p,b)
Formulation p b X
SDDP(1) 1SDDP(p) pSDDPX(p, SWE) p 1 SWESDDPX(p; SWE; Pw ) p 1 SWE; Pw
Table 3Average Annual Results-Differences With Respect to the SDDP(1) Model
First configuration Second configuration
Model
Spillagereduction
(m3=s)
Annual gainof energy
(GWh)
Spillagereduction
(m3=s)
Annual gainof energy
(GWh)
Net purchasesreduction
(%)
SDDP(p) 15 54 10 30 4.0SDDPX(p, SWE) 35 101 26 50 6.6SDDPX(p; SWE; Pw ) 33 105 28 51 6.7
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 7
79
storage levels in the upstream reservoirs tend to reduce the spillage losses throughout the system, there-fore increasing the total energy output and compensating for the reduced efficiency at the storage powerplants. With better forecasts throughout the system, the upstream reservoirs are operated at lower pool ele-vation to prevent excessive spills not only at the reservoir site, but also downstream.
Since the decisions taken during the drawdown phase have consequences beyond the summer, Figure 3also explores the statistical distribution of the storage decisions for weeks 14, 15, and 16, when storage lev-els are the lowest. In week 14, for example, when the simple SDDP(1) and SDDP(p) formulations are used,the volume in storage exceeds 2 km3 100% of the time. However, when the exogenous variables areincluded in the algorithm, storage volumes exceed this capacity less than 40% of the time, meaning that 6years out of 10, the storage levels in week 14 will be lower than 2 km3. These results clearly indicate thatthe incorporation of exogenous variables yields policies that better anticipates snowmelt runoff, thereforeavoiding massive spills.
The impact of these decisions on the amount of water spilled is further analyzed in Figure 4 where we cansee the probability distributions of annual spillages (left plots) at each power plant for two SDDP formula-tions: the simple SDDP(1) and the more sophisticated SDDPX(p; SWE; Pw). With SDDPX, the spills arereduced regardless of the hydrologic conditions, and the reduction is more pronounced in RLSJ, the down-stream reservoir supplying the two largest power stations accounting for about two-thirds of the installed
Figure 2. Weekly mean storage at reservoirs Passes-Dangereuses (RPD) and Lac-St-Jean (RLSJ). First configuration.
0 2000 4000 60000
0.2
0.4
0.6
0.8
1
week14
Non
-Exc
eede
nce
prob
abili
ty [-
]
Storage Hm30 2000 4000 6000
0
0.2
0.4
0.6
0.8
1
week15
Storage Hm30 2000 4000 6000
0
0.2
0.4
0.6
0.8
1
week16
Storage Hm3
SDDP(1) SDDP(p) SDDPX(p,SWE) SDDPX(p,Fs)
Figure 3. Statistical distribution of the accumulated storages for the largest reservoirs for weeks 14, 15, and 16. Firstconfiguration.
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 8
80
capacity. During dry years, however, the differences between the statistical distributions become marginalbecause the SDDP formulations with and without exogenous variables generate similar policies (from adecision-making point-of-view, the exogenous variables become informationless). For the most upstreamreservoir (RPD), the large storage to inflow ratio implies that this reservoir does not spill 3 years out of 4.When spills do occur, SDDPX release decisions significantly reduce the discharges through the spillway,therefore having a positive repercussion on downstream infrastructure: the smaller reservoir (RCD) and therun-of-river power plant (CCS).
The exogenous hydrologic state variables also have an impact on the marginal water values, which corre-spond to the Lagrange multipliers associated with the mass balance equation (4). Figure 4 (right plots) com-pares the statistical distributions of the marginal water values in the system with the SDDP(1) andSDDPX(p; SWE; Pw ) formulations only (for clarity, the two intermediate formulations are not shown). As wecan see, SDDPX-derived policies yield larger water values across a wide range of hydrologic conditions. At
35 40 45 50 550
0.5
1
Marginal water values $/1000 m3
18 20 22 24 260
0.5
1
Marginal water values $/1000 m3
14 16 18 200
0.5
1
Marginal water values $/1000 m3
10 11 12 13 140
0.5
1
Marginal water values $/1000 m3
7.5 8 8.5 9 9.50
0.5
1
Marginal water values $/1000 m3
0 10 20 30 400
0.5
1
Spillage losses m3/s
Reservoir Passes-Dangereuses
0 20 40 600
0.5
1
Spillage losses m3/s
Reservoir Chute du Diable
0 20 40 60 800
0.5
1
Spillage losses m3/sNon
-Exc
eede
nce
prob
abili
ty [-
]
Chute-a-la-Savane
0 200 400 6000
0.5
1
Spillage losses m3/s
Reservoir Lac-St-Jean
0 100 200 3000
0.5
1
Spillage losses m3/s
Shipshaw
SDDP(1) SDDPX(p,SWE,Pw)
Figure 4. Statistical distributions of (left) spillage losses and (right) marginal water values. First configuration.
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 9
81
any given power plant, reservoir operators would therefore be willing to pay more for the same unit ofwater under SDDPX release policies because, as explained above, unproductive spills are reduced.
3.2. Joint Optimization of Physical and Financial AssetsLet us now move to the second configuration. Incorporating the secondary operating objectives (recreationand flood control in Lac-Saint-Jean Reservoir) and the load commitment reveals one main change withrespect to the first configuration: the gains in energy generation that we can expect with the SDDP(p) andSDDPX formulations are reduced (Table 3). With the most complex SDDPX formulation, the average annualgain is about 105 GWh in the first configuration. In the second configuration, the annual gain is reduced to51 GWh. It turns out that by constraining the system and forcing the operators to buy increasingly moreexpensive energy to meet the load, the resulting SDDPX decisions tend to hedge more against the hydro-logical risk: the operating policies tend to store more water, accepting immediate, shorter, energy deficitsto reduce the probability of greater, increasingly more costly, energy shortages in the future. Figure 5 showsthe average weekly storages associated to the most complex SDDPX formulation for both configurations.As discussed earlier, the higher storage levels in the second, more restrictive, configuration tend to increasethe unproductive spills throughout the system, therefore affecting the total energy output.
Compared to SDDP(1), the reduction in power output is however less important with SDDPX-derived poli-cies; the net energy purchases (i.e., differences between purchases and sales) are reduced by 6.6% and 6.7%when the exogenous variables are included (Table 3).
The results can also be analyzed in terms of the annual power efficiency of the system, which is definedhere as the ratio between the power produced by the system and the total outflow. Generally speaking, theaverage efficiency is improved when more hydrologic variables are incorporated in the state space vector(see Table 4). We can see that the power efficiency of the system actually increases when the storage levels
in the head reservoir are lowered. This apparently counter-intuitiveresult is due to the characteristics of the RT cascade where a signifi-cant portion of the energy is generated by downstream run-of-riverpower plants that are prone to large spills especially if the head reser-voir is also spilling. In a cascade with storage power plants, the conclu-sions would probably be different: power efficiency would increasewith the storage levels.
Figure 6 shows, for both configurations, the statistical distributions ofthe annual differences in power efficiency between SDDPX(p; SWE; Pw)and SDDP(1) formulations. As we can see, the system’s efficiency is
4 8 12 16 20 24 28 32 36 40 44 48 522000
3000
4000
5000
6000
7000
8000
9000
10000
11000
Weeks
Sto
rage
[Hm
3 ]
SDDPX(p,SWE,Pw) 1st Config
SDDPX(p,SWE,Pw) 2nd Config
Figure 5. Average weekly storages at reservoirs Passes-Dangereuses (RPD) and Lac-St-Jean (RLSJ)-FormulationSDDPX(p,SWE,Pw) for both configurations.
Table 4Average Power Efficiency of the System
Average efficiency (MW/m3 s21)
First configuration Second configuration
SDDP(1) 1.3976 1.3997SDDP(p) 1.4019 1.4014SDDPX(p, SWE) 1.4056 1.4033SDDPX(p; SWE; Pw ) 1.4060 1.4034
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 10
82
enhanced more than 65% of the time (the nonexceedance probability > 0.35) when the exogenous hydro-logic variables are taken into account. During the rest of the time, failed inflow forecasts lead to ill-informeddecisions which affect the overall efficiency of the system in the following months. Of interest is the factthat the variability of the efficiency gains/losses is less pronounced for the second configuration. In thatcase, both the upside and the downside of the hydrological risk associated with imperfect forecasts arepartly suppressed. As energy shortages are increasingly costly, i.e., compensated by purchases throughincreasingly more expensive contracts, the SDDPX release policies acts as a hedging mechanism therebyyielding more conservative release decisions over the entire spectrum of the hydrologic uncertainty.
4. Discussion and Conclusions
As the availability of various hydroclimatic information keeps increasing due to advances in environmentalmonitoring systems, there is a need for decision-making processes and tools that can process this information.SDDP has for many years been the most advantageous model for optimizing in a stochastic framework theoperating of large multiple-reservoir systems. We present a variant of the traditional SDDP algorithm in whichdifferent exogenous hydrologic variables such as snow water equivalent and/or sea surface temperature canbe included in the state space vector together with the past inflows. The incorporation of these exogenous var-iables relies on a built-in MPARX model to generate the inflows during both phases of the algorithm (backwardoptimization and forward simulation). The results are consistent with previous studies and show that goodinformation about future flows can result in more efficient hydropower system operations. This paper demon-strates that the SDDP model can show operators how to achieve such gains in large hydropower systems.
Appendix A: SDDPX Parameter Estimation
The appendix shows how to analytically derive the parameters of the hyperplanes used to approximate thefuture benefit function in SDDPX. As indicated in section 2.2, the main modification lies in the calculation ofthe hyperplanes’ parameters ut11; vt11; bt11, and Ct11 (see equation (5)). In particular, Ct11 which can bewritten as:
Ct11ht115ct11;1qt1ct11;2qt211 . . . 1ct11;pqðt2pÞ11
1ct11;p1jXðt2jÞ111 . . . 1ct11;p1bXðt2bÞ11
(A1)
According to the Karush-Kuhn-Tucker conditions for optimality, the derivative of the objective function withrespect to the state variables Si is given by:
-4 -2 0 2 4 60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
% Difference in Power Efficiency
Non
-Exc
eeda
nce
prob
abili
ty [-
]SDDP(1)
SDDPX(p,SWE,Pw) 1st config
SDDPX(p,SWE,Pw) 2nd config
Figure 6. % of difference in the power efficiency with respect to SDDP(1) formulation for both configurations.
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 11
83
@F@Si
52X
ki@gi
@Si(A2)
where ki is the dual information (Lagrange multiplier) of the convex optimization problem and gi is the con-straints (Kuhn & Tucker, 1951). Then, for the specific case at stage t and using constraints (4), (5), (8), and(10), the change of the one-stage objective function Ft respect to the state variables st; wt; ht , can be deter-mined by:
@Ft
@st5kw;t1
XH
h51
khhp;tw
ht11=2 (A3)
@Ft
@wt5ke;t (A4)
@Ft
@ht5kw;t
@qt
@ht1XL
l51
klc;t
@ Clt11ht11
� �@ht
(A5)
where kw;t; ke;t; klc;t , and kh
hp;t are, respectively, the vectors with the dual information associated to the massbalance (4), energy balance (10), the L cuts of the benefit-to-go function (5), and the H linear segments ofthe power functions (8).
Using (A1), we can calculate the derivatives in the above equation (A5) for each hydrologic variable½qt21; qt22,. . .,qt2p; Xt21; Xt2j,. . ., Xt2b�:
@ Ct11ht11ð Þqt21
5clt11;1
@qt
@qt211cl
t11;2
@ Ct11ht11ð Þqt22
5clt11;1
@qt
@qt221cl
t11;3
�
@ Ct11ht11ð Þqt2p
5clt11;1
@qt
@qt2p
@ Ct11ht11ð ÞXt2j
5clt11;1
@qt
@Xt2j1cl
t11;p1j11
�
@ Ct11ht11ð ÞXt2b
5clt11;1
@qt
@Xt2b
(A6)
Now, let us say that at stage t, s�
t ; w�
t , and h�
t 5 ½q�t21; q�
t22,. . ., q�
t2p; X�
t2j,. . ., X�
t2b� are sampled and, in orderto include the stochasticity of the problem, K vectors of inflows qK
t are generated using the MPARX(p,b)(equation (12)). Since Fk
t , which will be added to the expected-benefit-to-go function at stage ðt21Þ, can beapproximated by:
Fkt � ul;k
t s�t 1vl;k
t w�t 1Cl;k
t h�t 1bl;k
t (A7)
the slopes ul;kt ; v
l;kt ;C
l;kt 5½cl;k
t;1; . . . ; cl;kt;p; c
l;kt;p1j; . . . ; cl;k
t;p1b� are determined for each hydrologic scenario k usingequations (A3), (A4), and (A6):
@Fkt
@stðjÞ5ul;k
t ðjÞ5kkw;tðjÞ1
XH
h51
kh;khp;tðjÞw
ht11ðjÞ=2 (A8)
@Fkt
@wtðcÞ5vl;k
t ðcÞ5kke;tðcÞ (A9)
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 12
84
@FKt
@qt21ðjÞ5cl;k
t;1ðjÞ5 kl;kw;tðjÞ1
XL
l51
kl;kc;tðjÞcl
t11;1ðjÞ !
@qt
@qt211XL
l51
kl;kc;tðjÞcl
t11;2ðjÞ
@FKt
@qt22ðjÞ5cl;k
t;2ðjÞ5 kl;kw;tðjÞ1
XL
l51
kl;kc;tðjÞcl
t11;1ðjÞ !
@qt
@qt221XL
l51
kl;kc;tðjÞcl
t11;3ðjÞ
�
@FKt
@qt2pðjÞ5cl;k
t;pðjÞ5 kl;kw;tðjÞ1
XL
l51
kl;kc;tðjÞcl
t11;1ðjÞ !
@qt
@qt2p
@FKt
@Xt2jðjÞ5cl;k
t;p1jðjÞ5 kl;kw;tðjÞ1
XL
l51
kl;kc;tðjÞcl
t11;1ðjÞ !
@qt
@Xt2j1XL
l51
kl;kc;tðjÞcl
t11;p1j11ðjÞ
�
@FKt
@Xt2b5cl;k
t;p1bðjÞ5 kl;kw;tðjÞ1
XL
l51
kl;kc;tðjÞcl
t11;1ðjÞ !
@qt
@Xt2b
(A10)
Defining gðjÞ as:
gtðjÞ5kl;kw;tðjÞ1
XL
l51
kl;kc;tðjÞcl
t11;1ðjÞ (A11)
and by using (12) to find the derivatives of qt respect to the hydrologic variables, the set of equation (A10)can be rewritten as:
cl;kt;1ðjÞ5gtðjÞ
rqt ðjÞrqt21ðjÞ
/t;1ðjÞ1XL
l51
kl;kc;tðjÞcl
t11;2ðjÞ
cl;kt;2ðjÞ5gtðjÞ
rqt ðjÞrqt22ðjÞ
/t;2ðjÞ1XL
l51
kl;kc;t11ðjÞcl
t11;3ðjÞ
cl;kt;pðjÞ5gtðjÞ
rqt ðjÞrqt2pðjÞ
/t;pðjÞ
�
cl;kt;p1jðjÞ5gtðjÞ
rXt ðjÞrXt2jðjÞ
#t;jðjÞ1XL
l51
kl;kc;t11ðjÞcl
t11;p1j11ðjÞ
cl;kt;p1bðjÞ5gtðjÞ
rXt ðjÞrXt2pðjÞ
#t;bðjÞ
(A12)
Taking the expectation over the K artificially generated flows, the vector of slopes ult; vl
t;1; clt;1;
clt;2,. . .,cl
t;p; clt;p1j,. . .,cl
t;p1b can be determined:
ultðjÞ5
1K
XK
k51
ul;kt ðjÞ (A13)
vltðcÞ5
1K
XK
k51
vl;kt ðcÞ (A14)
clt;arxðjÞ5
1K
XK
k51
cl;kt;arxðjÞ; (A15)
8arx51; 2; . . . ; p; p1j; . . . ; p1b
Finally, the constant term is given by:
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 13
85
blt5
1K
XK
k51
Fkt 2X
J
ultðjÞs
�
t ðjÞ2X
C
vltðcÞw
�
t ðcÞ . . .
2X
J
clt;1ðjÞq
�t21ðjÞ2
XJ
clt;2ðjÞq
�t22ðjÞ2
XJ
clt;pðjÞq
�t2pðjÞ . . .
2X
J
clt;p1jðjÞX
�
t2jðjÞ2 . . . 2X
J
clt;p1bðjÞX
�
t2bðjÞ
(A16)
ReferencesAhmad, A., El-Shafie, A., Razali, S. F. M., & Mohamad, Z. S. (2014). Reservoir optimization in water resources: A review. Water Resource Man-
agement, 28(11), 3391–3405. https://doi.org/10.1007/s11269-014-0700-5Archibald, T. W., Buchanan, C. S., McKinnon, K. I. M., & Thomas, L. C. (1999). Nested Benders decomposition and dynamic programming for
reservoir optimisation. Journal of the Operational Research Society, 50(5), 468–479. https://doi.org/10.1057/palgrave.jors.2600727Bellman, R. (1957). Dynamic programming (1 ed.). Princeton, NJ: Princeton University Press.Bras, R., Buchanan, R., & Curry, K. (1983). Real time adaptive closed loop control of reservoirs with the High Aswan Dam as a case study.
Water Resources Research, 19(1), 33–52. https://doi.org/10.1029/WR019i001p00033Cerisola, S., Latorre, J. M., & Ramos, A. (2012). Stochastic dual dynamic programming applied to nonconvex hydrothermal models. European
Journal of Operational Research, 218(3), 687–697. https://doi.org/10.1016/j.ejor.2011.11.040Cot�e, P., Haguma, D., Leconte, R., & Krau, S. (2011). Stochastic optimisation of Hydro Quebec hydropower installations: A statistical compari-
son between SDP and SSDP methods. Canadian Journal of Civil Engineering, 1434(April), 1427–1434. https://doi.org/10.1139/l11-101Desreumaux, Q., Cot�e, P., & Leconte, R. (2014). Role of hydrologic information in stochastic dynamic programming: A case study of the Kemano
hydropower system in British Columbia. Canadian Journal of Civil Engineering, 41(9), 839–844. https://doi.org/10.1139/cjce-2013-0370Faber, B., & Stedinger, J. (2001). Reservoir optimization using sampling SDP with ensemble streamflow prediction (ESP) forecasts. Journal of
Hydrology, 249, 1113–1133. https://doi.org/10.1016/S0022-1694(01)00419-XGelati, E., Madsen, H., & Rosbjerg, D. (2013). Reservoir operation using El Ni~no forecasts case study of Daule Peripa and Baba, Ecuador.
Hydrological Sciences Journal, 59(8), 1559–1581. https://doi.org/10.1080/02626667.2013.831978Gjelsvik, A., Mo, B., & Haugstad, A. (2010). Long and medium term operations planning and stochastic modelling in hydro-dominated
power systems based on stochastic dual dynamic programming. In S. Rebennack et al. (Eds.), Handbook of power systems I, Energy sys-tems (pp. 33–55). Berlin, Heidelberg: Springer. https://doi.org/10.1007/978-3-642-02493-1
Goor, Q., Kelman, R., & Tilmant, A. (2011). Optimal multipurpose multireservoir operation model with variable productivity of hydropowerplants. Journal of Water Resources Planning and Management, 137(3), 258–267. https://doi.org/10.1061/(ASCE)WR.1943-5452
Hamlet, A. F., & Lettenmaier, D. P. (1999). Columbia River streamflow forecasting based on ENSO and PDO climate signals. Journal of WaterResources Planning and Management, 125(6), 333–341.
Homen-de Mello, T., de Matos, V. L., & Finardi, E. C. (2011). Sampling strategies and stopping criteria for stochastic dual dynamic program-ming: A case study in long term hydrothermal scheduling. Energy Systems, 2(1), 1–31. https://doi.org/10.1007/s12667-011-0024-y
Karamouz, M., & Vasiliadis, H. V. (1992). Bayesian stochastic optimization of reservoir operation using uncertain forecasts. Water ResourcesResearch, 28(5), 1221–1232. https://doi.org/10.1029/92WR00103
Kelman, J., Stedinger, J., Cooper, L., Hsu, E., & Yuan, S. (1990). Sampling stochastic dynamic programming applied to reservoir operation.Water Resources Research, 26(3), 447–454. https://doi.org/10.1029/WR026i003p00447
Kim, Y., & Palmer, R. (1997). Value of seasonal flow forecasts in Bayesian stochastic programming. Journal of Water Resources Planning andManagement, 123(6), 327–335. https://doi.org/10.1061/(ASCE)0733-9496(1997)123:6(327)
Kim, Y-O., Eum, H., Lee, E., & Ko, I. (2007). Optimizing operational policies of a Korean multireservoir system using sampling stochasticdynamic programming with ensemble streamflow prediction. Journal of Water Resources Planning and Management, 133(1), 4–14.
Kristiansen, T. (2004). Financial risk management in the electric power industry using stochastic optimization. Advanced Modelling and Opti-mization, 6(2), 17–24.
Kuhn, H. W., & Tucker, A. W. (1951). Nonlinear programming. In Proceedings of the second Berkeley symposium on mathematical statisticsand probability (pp. 481–492). Berkeley, CA: University of California Press.
Kwon, H.-H., Brown, C., Xu, K., & Lall, U. (2009). Seasonal and annual maximum streamflow forecasting using climate information: Applicationto the Three Gorges Dam in the Yangtze River basin, Hydrological Sciences Journal, 54(3), 582–595. https://doi.org/10.1623/hysj.54.3.582
Labadie, J. W. (2004). Optimal operation of multireservoir systems: State-of-the-art review. Journal of Water Resources Planning and Man-agement, 130(2), 93–111. https://doi.org/10.1061/(ASCE)0733-9496(2004)130:2(93)
Ljung, L. (1999). System identification: Theory for the user. Englewood Cliffs, NJ: Prentice-Hall.Lohmann, T., Hering, A. S., & Rebennack, S. (2015). Spatio-temporal hydro forecasting of multireservoir inflows for hydro-thermal schedul-
ing. European Journal of Operational Research, 255(1), 243–258. https://doi.org/10.1016/j.ejor.2016.05.011Macian-Sorribes, H., Tilmant, A., & Pulido-Velazquez, M. (2016). Improving operating policies of large-scale surface-groundwater systems
through stochastic programming. Water Resources Research, 53, 1407–1423. https://doi.org/10.1002/2016WR019573Matos, V. L. D., & Finardi, E. C. (2012). A computational study of a stochastic optimization model for long term hydrothermal scheduling.
International Journal of Electrical Power & Energy Systems, 43(1), 1443–1452. https://doi.org/10.1016/j.ijepes.2012.06.021Mo, B., Gjelsvik, A., & Grundt, A. (2001). Integrated risk management of hydro power scheduling and contract management. IEEE Transac-
tions on Power Systems, 16(2), 216–221. https://doi.org/10.1109/59.918289New York Independent System Operator (NYISO) (2017). Hourly custom report-Time-Weighted/Integrated Real-Time LBMP-Reference Bus.
Rensselaer, NY: New York Independent System Operator. Retrieved fromhttp://www.nyiso.comOliveira, R., & Loucks, D. P. (1997). Operating rules for multireservoir systems. Water Resources Research, 33(4), 839–852. https://doi.org/10.
1029/96WR03745Pereira, M. (1989). Optimal stochastic operations scheduling of large hydroelectric systems. International Journal of Electrical Power &
Energy Systems, 11(3), 161–169. https://doi.org/10.1016/0142-0615(89)90025-2Pereira, M. V. F., & Pinto, L. M. V. G. (1991). Multi-stage stochastic optimization applied to energy planning. Mathematical Programming,
52(1–3), 359–375. https://doi.org/10.1007/BF01582895
AcknowledgmentsAuthors acknowledge the CRDPJproject funded by NSERC and Rio TintoAlcan, and NSERC Strategic NetworkProject FloodNet. We also thank thereviewers for their constructivecomments and suggestions. We thankDr. Charles Roug�e and Dr. Sara S�eguinfor their valuable discussions.Supporting data are not publiclyavailable and belong to Rio Tinto, theprivate company operating the powersystem used to illustrate the proposedmethodology. These data can berequested by contacting the thirdauthor ([email protected]).
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 14
86
Pereira-Cardenal, S. J., Mo, B., Gjelsvik, A., Riegels, N. D., Arnbjerg-Nielsen, K., & Bauer-Gottwein, P. (2016). Joint optimization of regionalwater-power systems. Advances in Water Resources, 92, 200–207. https://doi.org/10.1016/j.advwatres.2016.04.004
Pina, J., Tilmant, A., & Anctil, F. (2016). Horizontal approach to assess the impact of climate change on water resources systems. Journal ofWater Resources Planning and Management, 143, 04016081-1. https://doi.org/10.1061/(ASCE)WR.1943-5452.0000737
Poorsepahy-Samian, H., Espanmanesh, V., & Zahraie, B. (2016). Improved inflow modeling in stochastic dual dynamic programming. Journalof Water Resources Planning and Management, 142, 04016065-1. https://doi.org/10.1061/(ASCE)WR.1943-5452.0000713
Pritchard, G. (2015). Stochastic inflow modeling for hydropower scheduling problems. European Journal of Operational Research, 246, 496–504. https://doi.org/10.1016/j.ejor.2015.05.022
Rani, D., & Moreira, M. M. (2009). Simulation optimization modeling: A survey and potential application in reservoir systems operation.Water Resource Management, 24(6), 1107–1138. https://doi.org/10.1007/s11269-009-9488-0
Raso, L., Malaterre, P. O., & Bader, J. C. (2017). Effective streamflow process modeling for optimal reservoir operation using stochastic dualdynamic programming. Journal of Water Resources Planning and Management, 143, 04017003-1. https://doi.org/10.1061/(ASCE)WR.1943-5452.0000746
Rotting, T., & Gjelsvik, A. (1992). Stochastic dual dynamic programming for seasonal scheduling in the Norwegian power system. IEEE Trans-actions on Power Systems, 7(1), 273–279. https://doi.org/10.1109/59.141714
Shapiro, A., Tekaya, W., da Costa, J. P., & Soares, M. P. (2013). Risk neutral and risk averse Stochastic Dual Dynamic Programming method.European Journal of Operational Research, 224(2), 375–391. https://doi.org/10.1016/j.ejor.2012.08.022
Stedinger, J. R., Sule, B. F., & Loucks, D. P. (1984). Stochastic dynamic programming models for reservoir operation optimization. WaterResources Research, 20(11), 1499–1505. https://doi.org/10.1029/WR020i011p01499
Tejada-Guibert, J. A., Johnson, S. A., & Stedinger, J. R. (1995). The value of hydrologic information in stochastic dynamic programming mod-els of a multireservoir system. Water Resources Research, 31(10), 2571–2579. https://doi.org/10.1029/95WR02172
Tilmant, A., & Kelman, R. (2007). A stochastic approach to analyze trade-offs and risks associated with large-scale water resources systems.Water Resources Research, 43, W06425. https://doi.org/10.1029/2006WR005094
Tilmant, A., & Kinzelbach, W. (2012). The cost of noncooperation in international river basins. Water Resources Research, 48, W01503.https://doi.org/10.1029/2011WR011034
Tilmant, A., Pinte, D., & Goor, Q. (2008). Assessing marginal water values in multipurpose multireservoir systems via stochastic program-ming. Water Resources Research, 44, W12431. https://doi.org/10.1029/2008WR007024
Yeh, W. W.-G. (1985). Reservoir management and operations models: A state-of-the-art review. Water Resources Research, 21(12),1797–1818. https://doi.org/10.1029/WR021i012p01797
Zahraie, B., & Karamouz, M. (2004). Hydropower reservoirs operation: A time decomposition approach. Scientia Iranica, 11(1–2), 92–103.Zhao, T., Zhao, J., & Yang, D. (2014). Improved dynamic programming for hydropower reservoir operation. Journal of Water Resources Plan-
ning and Management, 140(March), 365–374. https://doi.org/10.1061/(ASCE)WR1943-5452.0000343
Water Resources Research 10.1002/2017WR021701
PINA ET AL. SDDPX EXOGENOUS VARIABLES 15
87
Appendix C
Autoregressive modeling
C.1 Multisite periodic autoregressive model MPAR
Autoregressive (AR) models have been extensively used in hydrology and water resources since
1960’s. Its popularity and attractiveness rely on the simplest formulation and its intuitive type of
time dependence, variables values at time t are dependent to the preceding values. First introduce by
Thomas and Fiering (1962) and later by Box and Jenkins (1970). AR models can be represented as
models with constant parameters, parameters varying with time and combination of both. The models
with constant parameters are often implemented for modeling time series on the annual basis. Models
with periodic parameters are usually used with time series of intervals that are fraction of the year
(e.g. seasons, months, weeks, etc.) Salas et al (1980). The latter models, are referred as periodic AR,
PAR(p) models, and the periodicity may be in the mean, variance and/or autoregressive parameters.
Finally, since analysis in water systems involves time series at various geographic locations mod-
els which evaluate the spatial correlation are required, then let us introduce the multivariate spacial
models Multisite PAR, MPAR.
The multi site ( j) periodic process can be modeled by an autoregressive model of order p represented
by:
(qt( j)−µqt ( j)
σqt ( j)
)=
p
∑i=1
φi,t( j)
(qt−i( j)−µqt−i( j)
σqt−i( j)
)+ εt( j) (C.1)
where qt is the time dependent variable for year v and time t, with t=1,2,...,52 weeks. µqt and σqt are
the periodic mean and standard deviation of qt , respectively, φi,t( j) are the autoregressive parameters
of the p order periodic model, and ξt is a time independent stochastic noise with mean zero and
periodic variance σ2ε,t( j).
88
C.1.1 Parameter estimation
Moments
The set of parameters to be determined in the model are: µt( j), σt( j), φi,t( j) and σ2ε,t( j). Given N
years of historical incremental reservoirs inflows data qv,t( j) available at site j, the parameters µqt ( j),
σqt ( j) can be estimated by the first and second moment as:
µqt ( j) = E[qt( j)
]≈ 1
N
N
∑v=1
qv,t( j) (C.2)
σ2qt( j) = E
[q2
t ( j)]≈ 1
N−1
N
∑v=1
(qv,t( j)−µt( j))2 (C.3)
defining the vector Zt( j) [N×1] of normalized incremental inflows as:
Zt( j) =
(qv,t( j)−µqt ( j)
σqt ( j)
), ∀v = 1 . . .N (C.4)
by using equation C.1 the normalized inflows can be expressed as:
Zt( j) = φ1,t( j)Zt−1( j)+ · · ·+φp,t( j)Zt−p( j)+ εt( j) (C.5)
The temporal Lag-k autocorrelation ρt(k) between qv,t and qv,t−k for each season t is obtained by
multiplying C.5 by Zt−k and taking the expectation term by term:
E[Z( j)Zt−k( j)] = φ1,t( j)E[Zt−1( j)Zt−k( j)]+ · · ·+φp,t( j)E[Zt−p( j)Zt−k( j)]+E[εt( j)Zt−k( j)] (C.6)
ρt (k) = φ1,tρt−1(|k−1|)+ · · ·+φp,tρt−p(|k−1|) (C.7)
Equation C.7 constitutes a set of p linear equations, for k=1,2,...,p
ρt (1) = φ1,tρt−1(0)+φ2,tρt−2(1)+ · · ·+φp,tρt−p(1−p)
ρt (2) = φ1,tρt−1(1)+φ2,tρt−2(0)+ · · ·+φp,tρt−p(2−p)
... (C.8)
ρt (p) = φt,1ρt−1(p−1)+φt,2ρt−2(p−2)+ · · ·+φt,pρt−p(0)
89
which may be written for each j as:
ρt (1)
ρt (2)...
ρt (p)
=
1 ρt−1(1) . . . ρt−1(p−1)
ρt−1(1) 1 . . . ρt−2(p−2)...
......
ρt−1(p−1) ρt−2(p−2) . . . 1
φ1,t
φ2,t...
φp,t
(C.9)
in a matrix form:
ρ t( j) =COV q,t( j) ·Φt( j) (C.10)
where COV q,t( j) is the periodic covariance and Φt( j) and ρ t( j) respectively, are the vectors of autore-
gressive coefficients of the MPAR(P) and autocorrelation coefficients. Then φi,t( j) can be calculated
as:
Φt( j) =COV q,t( j)−1 ·ρ t( j) (C.11)
According to Salas et al (1980) the residual variance for each t, σ2ε,t( j) can be estimated as a function
of the periodic autoregressive coefficients φi,t( j) and the periodic autocorrelation coefficients ρi,t( j).
σ2ε,t( j) = 1−
p
∑j=1
φ j,t( j)ρ j,t( j) (C.12)
Least square parameter estimation
The least square estimation method is based on finding the estimates φi,t( j) so that the sum of the
squared differences between the observed Zt( j) values and the expected values Zt( j)= φ1,t( j)Zt−1( j)+
· · ·+ φp,t( j)Zt−p( j), is minimized.
minφt
{∑(Zt( j)− Zt( j))2
}(C.13)
Let us first say that Zt can be expressed in the matrix form
Zt( j) = Xt( j)Φt( j) (C.14)
90
where Xt( j) contains [Zt−1( j), . . . ,Zt−p( j)] and Φt( j) = [φ1,t( j), . . . , φp,t( j)] then from equation C.13
the objective function can be written as:
Θ( j) = ∑(Zt( j)−Xt( j)Φt( j))2
= ∑εt( j)2 = εᵀt ( j)εt( j)
= [Zt( j)−Xt( j)Φt( j)]ᵀ[Zt( j)−Xt( j)Φt( j)] (C.15)
To find the minimum, all partial derivatives of the sum respect to the estimates must be equal to zero:
∂Θ
∂Φt=
∂
∂Φt[Zt −XtΦt ]
ᵀ[Zt −XtΦt ] =−Xtᵀ[Zt −XtΦt ] = 0 (C.16)
Φt( j) = [Xtᵀ( j)Xt( j)]−1Xt
ᵀ( j)Zt( j) (C.17)
which represents the same correlation cross-correlation structure identified in equation C.11.φ1,t
φ2,t...
φp,t
=
∑Zt−1
2∑Zt−1Zt−2 . . . ∑Zt−1Zt−p
∑Zt−1Zt−2 ∑Zt−22 . . .
......
...
∑Zt−1Zt−p . . . ∑Zt−p2
−1
∑Zt−1Zt
∑Zt−2Zt...
∑Zt−pZt
(C.18)
Moreover, the Hessian matrix ∂ 2Θ
∂Φ2 must be positive definite
∂ 2Θ
∂Φ2 = Xtᵀ( j)Xt( j) positive definite (C.19)
The unbiased estimate of the variance of the model can be constructed as:
σ2ε,t( j) =
1N− p
εᵀ( j)ε( j) (C.20)
C.1.2 Stochastic noise
Assuming that the noise εt( j) follows a 3-parameters (µv( j), σv( j) and κt( j)) log normal distribution:
fεt( j) =1
(εt( j)−κt( j))√
2πσv( j)e−0.5
(log(εt ( j)−κt ( j))−µv( j)
σv( j)
)2
(C.21)
91
with mean µεt and variance σ2ε,t
µεt ( j) = κt( j)+ e
(µv( j)+ σ2
v ( j)2
)
σ2ε,t( j) = e2(µv( j)+σ2
v ( j))+ e(2µv( j)+σ2v ( j)) (C.22)
to ensure non-negative inflows qt( j)> 0 from equation C.1 the lower bound κt( j) is defined as:
εt( j)>−µqt ( j)σqt ( j)
−p
∑i=1
φi,t( j)
(qt−i( j)−µqt−i( j)
σqt−i( j)
)= κt( j) (C.23)
and the parameters µv( j) and σv( j) are determined as:
µv( j) = 0.5Log σ2ε,t
Λ( j)(Λ( j)−1) (C.24)
σv( j) =√
log(Λ( j)) (C.25)
Λt( j) = 1+ σ2ε,t
κ2t ( j) (C.26)
Finally the standardized stochastic noise is estimated as:
Vt( j) =log(εt( j)−κt( j))−µv( j)
σv( j)(C.27)
C.1.3 Spatial cross correlation
The spatial statistical dependence of reservoir inflows is introduced by the lag0-covariances and cross
variances of the standardized stochastic noise. Defining the vector Vt as the collection of independent
standardized noises of each j node of the system Vt = [Vt(1), . . .Vt(J)], the spatial model can be written
as:
Vt = AtWt (C.28)
where Wt is a column vector of J independent elements consisting of white noises, normally distributed
with zero mean and variance equal to 1. The estimation of matrix A can be obtained from the Cholesky
factorization of the covariance matrix of standardized noise at each node j:
At Aᵀt =Cov(Vt) (C.29)
92
C.1.4 Example parameter estimation Gatineau River Basin
The set of parameters to be determined in the hydrologic model equation C.1 are: the mean µt( j),
standard deviationσt( j), autoregressive parameters φi,t( j) and the variance of the residuals σ2ε,t( j).
Given the N=28 years of weekly incremental inflows data qv,t( j) available at each site j (see: Table
3.1, Figure3.2). For each node j and each week the parameters are determined as follow.
Mean and standard deviation
Figure C.1 displays the parameters µqt ( j), σqt ( j) estimated by the first and second moment equation
C.2 and C.3. The nival regime can be described by a very high discharge during spring-summer after
the snow melt season (weeks 12-24). Great variability during spring and autumn and low flows during
winter.
4 8 12 16 20 24 28 32 36 40 44 48 520
10
20
30
40
50
60Node 1: Cabonga
[Hm
3]
4 8 12 16 20 24 28 32 36 40 44 48 520
50
100
150
200Node 3: Paugan
[Hm
3]
weeks
4 8 12 16 20 24 28 32 36 40 44 48 520
100
200
300
400
500Node 2: Baskatong
[Hm
3]
4 8 12 16 20 24 28 32 36 40 44 48 520
5
10
15
20
25
30Node 4: Chelsea
[Hm
3]
weeks
µq
t
σq
t
µq
t
σq
t
µq
t
σq
t
µq
t
σq
t
Figure C.1: Mean and standard deviation
Autoregressive parameters
As an example, let us take week 15 of the inflow series at the Baskatong reservoir and let us explore
the correlation structure with 12 preceding weeks. The autoregressive parameters of model for all p
orders are determined. To identify the order of the autoregressive model, four strategies are evaluated:
i) partial autocorrelation function, ii) the evolution of the mean square error, iii) parsimony of the
models and iv) the stationarity conditions
93
The partial autocorrelation function is another way of representing the time dependence structure, and
it is useful for helping to identify the type of order of the model. In the set of equations C.9 or C.17,
the partial autocorrelation is given by the last coefficient φp,t , and is practical zero beyond a particular
lag, if lies inside the 1−α probability limits Box and Jenkins (1970).[−u1− α
2√N
;u1− α
2√N
](C.30)
where uα is the quantile of the standard normal distribution α the probability level (α = 5%), and N
is the sample size.
The principle of parsimony in model building can be considered into account by using the mathemat-
ical formulation Akaike Information Criterion (AIC) or Bayesian information criterion (BIC). Under
this criterion the model which gives the minimum AIC and BIC is the one selected.
AIC = N · ln(σ2ε,t( j))+2p+1 (C.31)
BIC = N · ln(σ2ε,t( j))+ p · log(N) (C.32)
Figure C.2 displays for each order p, (1) the partial correlogram φk and the 95% confidence interval,
(2) the mean square error MSE, (3) AIC and (4) BIC. From the collection of figures we can perceived
that models of order p =2, p =3 and p =6 are good candidates to be selected. The partial correlogram
reveals strong correlation between inflows of week 15 and week 12 (p =3) and week 9 (p =6). Like-
wise, except for the order p =4, the evolution of the MSE displays the important reduction in the error
function while increasing the order of the autoregressive model until p =6. Finally, the mathematical
information criterion AIC and BIC, displays local minimum values for model of order p =2 and p =6.
One further check on the parameters is to assure that σ2ε,t is always greater than zero. (Salas et al,
1980).
Stationary conditions
The stationary conditions are satisfied if the roots of the characteristic equation C.33 lie inside the unit
circle. That is, |ui|< 1, i= 1,...,p.
up−φ1,t( j)up−1−φ2,t( j)up−2−·· ·−φp,t( j) = 0 (C.33)
If we isolate the autoregressive parameters and we evaluate the stationary conditions by calculating
the roots of the polynomial characteristic equation C.33. From table C.1 one can see that the first root
of the model of order p =6 is outside the unite circle∣∣uy∣∣> 1 .
94
1 2 3 4 5 6 7 8 9 10 11 12-0.5
0
0.5
1
PACF
( k)
Order(p)
(1)
1 2 3 4 5 6 7 8 9 10 11 120.2
0.3
0.4
0.5
0.6
0.7
Order(p)
MSE
(2)
1 2 3 4 5 6 7 8 9 10 11 12-15
-10
-5
0
5
10
Order(p)
AIC
(3)
1 2 3 4 5 6 7 8 9 10 11 12-10-505
10152025
Order(p)BIC
(4)
Figure C.2: (1) Partial correlogram PACF (2) Mean square error MSE (3) AIC (4) BIC
Table C.1: Autoregressive parameters φp,t and polynomial characteristic roots ui
p φp,t(2) roots ui
p =1 0.5643 0.5643
p =3 0.5289 0.9293-0.1212 -0.2002 +0.6731i
0.4583 -0.2002 - 0.6731i
p =6 0.4585 1.0330-0.0423 0.5289+ 0.6850i
0.3524 0.5289- 0.6850i
-0.0595 -0.4000+0.8875i
-0.2044 -0.4000-0.8875i
0.6102 -0.8323
Hydrologic modeling
Once the model is parametrized, it can be used to generate different hydrologic scenarios. For example
in Figure C3 and C4 we can see the average and standard deviation of 25 synthetically generated
inflow series at each node of the system using the MPAR(p) model. Generally speaking, we can see
that mean and standard deviation are pretty much conserved.
95
4 8 12 16 20 24 28 32 36 40 44 48 520
100
200
300
400
500
Node 2: BaskatongAverage
[hm
3]
Historical
Simulated
4 8 12 16 20 24 28 32 36 40 44 48 520
50
100
150
200
250Standard deviation
[hm
3]
weeks
Historical
Simulated
4 8 12 16 20 24 28 32 36 40 44 48 5210
20
30
40
50
60
Node 1: CabongaAverage
[hm
3]
Historical
Simulated
4 8 12 16 20 24 28 32 36 40 44 48 520
5
10
15
20Standard deviation
[hm
3]
weeks
Historical
Simulated
Figure C.3: Cabonga and Baskatong
4 8 12 16 20 24 28 32 36 40 44 48 520
50
100
150
200
Node 3: PauganAverage
[hm
3]
Historical
Simulated
4 8 12 16 20 24 28 32 36 40 44 48 520
10
20
30
40
50
60
70Standard deviation
[hm
3]
weeks
Historical
Simulated
4 8 12 16 20 24 28 32 36 40 44 48 520
5
10
15
20
25
30
Node 4: ChelseaAverage
[hm
3]
Historical
Simulated
4 8 12 16 20 24 28 32 36 40 44 48 520
2
4
6
8
10Standard deviation
[hm
3]
weeks
Historical
Simulated
Figure C.4: Paugan and Chelsea
96