Upload
danglien
View
223
Download
0
Embed Size (px)
Citation preview
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 1
Dispatching active distribution networks through electrochemical storage
systems and demand side managementP r o f . Ma r io P a o lo ne a nd P r o f. C o lin N . J one s
D ist r ibut e d Ele c t r ica l Sy st e ms L a bo ra t or yA ut o ma t ic C o nt r o l L a bor a to r y 3
EP F L
F u t u r e E l e c t ri c P o w e r S y s te ms a n d t h e E n e r g y T r a n s i ti o nI n t e r n a ti o n al c o n fe re n ce
C h a m p é r y , S w i t ze rl an d , F e b . 5 2 0 1 7
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 2
Motivations
Definition of a balance group (from the Swiss grid code):
“A balance group is a virtual construct for the purposes of billing and accounting. [...]. Every distribution grid operator, trader, power producer, supplier and end consumer must belong to a balance group. The balance group manager (BGM) can conduct energy transactions with other balance group managers at home and abroad, offload energy from power stations or transfer energy to end consumers. To do this the BGM sends schedules within the planning phase to Swissgrid. On completion of the energy deliveries, Swissgrid balances all import/export schedules of the balance group [...] as well as all measured feed - ins and feed - outs per balance group (measuring values which Swissgrid receives from all grid operators) and, in the event of deviations, charges the purchased or sold energy to the BGM as balance energy. The BGM is responsible for ensuring that his or her balance group is as balanced as possible at all times.”
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 3
Motivations, cont’d
§ Achieving dispatched-by-design operation of traditionally
stochastic prosumption allows reducing grid reserve
requirements.
§ The dispatch plan is built to satisfy a local objective, such as
peak shaving, load levelling or minimization of the cost of
imported electricity.
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 4
The topology of a disp atchable feeder
The operation of a group of stochastic prosumers (generation + demand) is dispatched according to a profile established the day before operation (called dispatch plan) by controlling the real power injection of the battery.
Sources of flexibility:
§ flexible demand (part II)
§ physical energy storage storage systems (part I)
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 5
Formulation – A two stage process
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 6
Formulation – Day-ahead planning
The dispatch plan is a sequence at 5 minute resolution that denotes the power flow at the grid connection point that the feeder should follow.
It is the sum between prosumption point predictions and the so-called offset profile:
The latter is with the objective of restoring an adequate battery state-of-energy such that, during operation, enough up/down-flexibility is available to compensate the mismatch between presumption and realization.
Section III describes the experimental facility used to vali-date the proposed control strategy. Section IV presents anddiscusses the results from the experimental validation. Finally,Section V summarizes the outcomes of this work and proposesthe perspectives.
II. METHODS
A. Problem statement
We consider a group of prosumers, for which we would liketo smooth the consumption profile (load leveling) and dispatchtheir operation. As anticipated, the problem is formulatedaccording to a two-stage procedure: day-ahead and intra-dayphase. In the day-ahead stage, the objective is to determinethe dispatch plan, namely the power consumption profilethat the group prosumers is willing to follow during real-time operation. The dispatch plan is built as the sum of theforecasted power consumption profile, obtained through data-driven forecasting, and an offset profile. This latter quantity,which is obtained by solving a convex optimization problem,has the objective of generating a dispatch plan with minimumvariance, namely with minimum variation with respect to itsaverage value such that, during operation, the BESS willcharge (discharge) when the power profile exceeds (is below)the levelled profile and viceversa.
The intra-day operation consist in controlling the BESSactive power injection in order to track the dispatch plan,namely compensating for deviations between the dispatch planand actual consumption, which are likely to differ due to theoffset profile and to forecasting errors. This is accomplishedusing MPC, as illustrated in section II-C2.
B. Day-ahead problem
The objective is to build the dispatch plan, namely thepower consumption profile that the feeder should follow duringoperation, the day after. The dispatch plan
bP is defined as thesequence of N = 288 (i.e., the number of 5-minute intervals in24 hours) average power consumption values for the incomingday. The feeder dispatch plan is composed by the sum of theprosumers forecasted consumption profile bL
t
and the offset
profile Ft
:
bPt
= bLt
+ Ft
t = 1, . . . , N (1)
which are determined using the process illustrated in the nexttwo paragraphs.
1) Prosumers data-driven forecasting: The prosumers fore-casted consumption profile, denoted by bL, is produced througha nonparametric black-box method based on vector auto-regression. We assume that D daily sequences of 5 minutesaverage power consumption measurements are known fromhistorical data: these are denoted by L d 2 RN , d =0, . . . , D� 1. For any index d, are also known i) the calendarday-of-year, ii) whether the day corresponds to a working day
2In the problem formulation we do not consider the operational constraintsassociated to the grid. In other words, we assume that the battery power ratingresults in grid voltages and currents within operational bounds. This is thecase for stiff medium voltage grids.
or a holiday and iii) the mean global horizontal irradiance(GHI) during that day. The day for which the forecast profileis to be computed is said target day and is identified by d⇤.At first, a set ⌦ of indexes d that are representative scenariosof the target day is determined. ⌦ is identified by retainingfrom the complete dataset the indices of the daily sequenceswith characteristics more similar to those of the target daywith regard to three conditions:
• being a working day or a holiday;• being in the same period of the year;• having similar weather conditions.
This is done by identifying subsequent shrinking subsets ofindices, through the following heuristic procedure:
• a first subset ⌦00 is composed by selecting the indicesthat correspond to working days if the d⇤ is such and toholidays otherwise;
• a subset ⌦0 is then obtained by retaining from ⌦00 the p0
indices having day-of-year closer to the one of d⇤;• finally, the set ⌦ is obtained by retaining from ⌦0 the p
(with p < p0) indices corresponding to days with meanGHI closer to the one forecasted for d⇤. The GHI iscalculated from publicly available cloud coverage forecastdata for the Lausanne area and by means of the modeldescribed in [7].
The values for p0 and p are chosen equal to 10 and 5respectively.Summarizing, the set ⌦ is composed of p indexes correspond-ing to days which are i) of the same kind as the target day, ii)closest in time to the target-day and iii) closest in amount ofradiation to the GHI forecast for the target-day d⇤.
The sequence of point predictions for the day d⇤, denotedby bL0, . . . , bLN�1, is obtained by equally averaging the dailysequences identified by the indexes in ⌦:
bLi
=1
|⌦|X
d2⌦
L d
i
i = 0, . . . , N � 1, (2)
where L d
i
denotes the value at the discrete time interval i ofthe scenario L d and |⌦| is the cardinality of the set ⌦.
2) Dispatch plan offset profile: The objectives of the offset
profile are• altering the dispatch plan so that it is with mininum
variance;• making sure that an adequate level of charge is available
in the BESS to achieve dispatchability during intra-dayoperation.
We define the average daily power consumption value as:
bLavg
=1
N
N
X
t=1
bLt
. (3)
The offset profile F o =�
F o
1 , . . . , Fo
N
�
is determined by aconstrained optimization problem that minimizes the move-ment of the forecasted consumption sequence bL1, . . . , bLN
around its average daily value bLavg
. The optimization problemconstraints are:
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 7
Formulation – Offset profile problem (non convex)We seek for the smallest offset profile F such that the battery state-of-energy is within bounds in the scenarios with highest and lowest possible prosumption .
Note that this is a non convex problem due to the sign operators .
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 8
Formulation – Offset profile problem (convex)The previous problem can be formulated as a convex one by writing the sign operator as the sum of two mutually exclusive terms. We define:
which are used to rewrite the previous optimization problem. The cost function achieves to keep the positive and negative components mutually exclusive.
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 9
Formulation – The real-time control problem (MPC)The objective is to track the dispatch plan. Since it consists in accomplishing a certain energy throughput, we rely on MPC rather a conventional feedback control loop to determine the current evolution while respecting BESS operational constraints. MPC is actuated at 10 sec resolution on a 5 min shrinking horizon by plugging in short-term prosumption forecasts and open-loop predictions of the BESS operational constraints (voltage and current).
Two formulations are possible:1. determining the BESS power to accomplish the energy throughput subject to BESS
constraints. However, BESS constraints are nonlinear and nonconvex.2. Determining the BESS current to minimize the distance from the target energy
throughput while subject to linear voltage and current constraints. However the cost function:
is in the form q(r(x)). To be convex, it requires r(x) to be convex (it is) and q convex nondecreasing (it is not), thus it is nonconvex.
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 10
Formulation – The new (convex) MPCThe BESS energy throughput in the 5 minute interval is the integral over time of the product between BESS DC current, voltage and converter efficiency alpha:
The BESS voltage dynamic evolution depends on the charge/discharge current. It can be modelled by using a three-time-contant (TTC) model as a function of the initial BESS state x_kas the following linear relantioship.
which replaced in the first expression leads to:
The expression above is the sum of two linear expressions and a quadratic form in the current. It is therefore convex provided that ψ is SDP, which has been numerically proven for the adopted TTC model.
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 11
Formulation – The new (convex) MPCWe use the previous result to formulate a convex equivalency of the original MPC optimization problem. This consists in maximizing the current (linear cost function) subject to the energy throughput being less or equal to the target energy throughput e_k (convex inequality).
Once the current is known from the MPC, it is multiplied by the voltage to determine the real power set-point to finally submit to the BESS converter.
§ value of the prosumption set-point to match (from the dispatch plan)
§ expected average consumption with short-term point prediction
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 12
Formulation – Modelling of energy storage
We apply grey-box modelling on offline measurements of the DC voltage vs. current to identify (linear) system dynamics.
§ Measurements are from dedicated experiments where the BESS was excited by using a pseudo-random binary signal (PRBS, a two-state signal with random duration).
§ Since parameters are state-of-charge (SOC) dependent, the identification experiment is carried out for different BESS SOC intervals.
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 13
Formulation – Modelling of energy storage, cont’d
§ The model which captures the best early to middle-range time dynamics is a third order linear model, an extension of the well known TTC.
§ The autocorrelation function shows i.i.d. model residuals.
§ For the MPC, we apply model scheduling, namely we select the set of parameters corresponding to the current SOC and we assume that this does not vary in the control period.
BESS equivalent circuit (set of parameters’ values for each considered SOC range).
Model prediction errors are i.i.d. (50% SOC).
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 14
The EPFL experimental set up
§ Single measurement point at the GCP.§ 350 kW peak demand during winter.§ 95 kWp roof-top PV installation.
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 15
Parameter Value
Nominal Capacity 720 kVA/560 kWhGCP Voltage 20 kVDC Bus Voltage Range 600/800 VCell Technology(Anode/Cathode)
Lithium Titanate Oxide (LTO)Nichel Cobalt Alumnium Oxide (NCA)
Number of racks 9 in parallel
Number of modules per rack
15 in series
Cells configuration per module
20s3p
Total number of cells 8100
Cell nominal voltage 2.3 V (limits 1.7 to 2.7 V)
Cell nominal capacity 30 Ah (69 Wh)Round-trip efficiency(AC side)
94-96%
Round-trip efficiency(DC side)
97-99%
The EPFL experimental setup – The BESS sp ecs
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 16
Results – 14/01/2016, operation
• Prosumption worst-case scenarios (shaded band)
• Prosumption point predictions (dashed)
• Offset plan (black).
• Dispatch plan (gray)• Composite power realisation at
the GCP (dashed) • Prosumption realization without
the battery correction (black)
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 17
Results – 14/01/2016, BESS data
BESS state-of-charge, DC Current and DC voltage with respective limits.
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 18
Results
Dispatched operation -- 14 Jan 2016https://snapshot.raintank.io/dashboard/snapshot/PuW1Rf5d470Q0gsT7UNponM25bGDNTRA
Dispatched operation -- 13 Jan 2016https://snapshot.raintank.io/dashboard/snapshot/cDS4IDniZjRiePXvusnmQXOmMwpGLnR6
Dispatched operation + Peak Shaving -- 22/06/2016https://snapshot.raintank.io/dashboard/snapshot/LSF3bPxtWYDjHVu6siEr1VPb92EXNkd6
Dispatched Operation + Load Levelling -- 14/03/2016https://snapshot.raintank.io/dashboard/snapshot/4ztn800czpAzEFRzbGOmWc1A2pKeC9ab
Dispatched operation (continuos operation) -- 16 to 19/03/2016https://snapshot.raintank.io/dashboard/snapshot/TNbEgP7j1AWhaW7cEK1ZiK3tY1Or7P4U
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 19
Conclusions – Part I§ A bottom-up approach to tackle the challenge of increasing reserve requirements due to
integration of larger shares of renewables.
§ Suitable to operate in current vertically operated power systems.
§ Fully decentralized control mechanism with no coordination requirements: complexity is masked behind the commitment of the operator to follow the dispatch plan.
§ No pervasive monitoring/control infrastructure.
§ Inherently allows to achieve local grid operational objectives, like peak shaving or load levelling.
§ The framework is flexible to include the control of other resources (see part II).
§ Grid constraints are not considered. It relies on the fact that storage can be sited and sized offline in the planning phase to mitigate localized network issues, thus without need of incorporating (which comes at the cost of much higher complexity and uncertainty).
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 20
The topology of a disp atchable feeder
The operation of a group of stochastic prosumers (generation + demand) is dispatched according to a profile established the day before operation (called dispatch plan) by controlling the real power injection of the battery.
Sources of flexibility:
§ flexible demand (part II)
§ physical energy storage storage systems (part I)
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 21
Outline – Part II
1. Dispatch plans for occupied buildings
2. Real-time control for dispatchability
3. Early experimental results (one week)
4. Optimal sizing
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 22
Dispatch Planning – Two Competing Objectives
❶ Minimize energy
❷ Maximize flexibility
Goal: Choose dispatch plan to maximize controllability during highly uncertain periods
Temp
eratur
eTe
mpera
ture
Cost
Low
High
Comfort
Low
High
Flexibility
Low
High
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 23
Optimal Dispatch Problem
Comfort metric
Day-ahead plan for the thermal (x, e) trajectory of the building and electrical dispatch d
Dispatch error
Energy cost
Thermal trajectory must be input-admissible (feasible) and comfortable C(w) for all likely
weather scenarios W
minx,u
�E
�costi ·ei + �(xi � x̄) + �ei � (di + pi)�2
�
s.t. (x, e) � C(w) �w � W
Building dispatch plan
Prosumer forecast error
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 24
Optimal Dispatch Problem
Uncertainty model• Weather• Occupancy• Prosumer consumption (PV & buildings)
Model of building and HVAC equipment
minx,u
�E
�costi ·ei + �(xi � x̄) + �ei � (di + pi)�2
�
s.t. (x, e) � C(w) �w � W
Day-ahead plan for the thermal (x, e) trajectory of the building and electrical dispatch d
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 25
Linear time-varying model and constraints
Building Modeling Tool http://la.epfl.ch/openbuild
ui thermal energy to each building zonexi thermal stateei energy consumed
Building Information Model Measured Building Data
openBuild
Building Geometry and Time of Year / Day
xi+1 = Aixi + Biui + Tiwi
(xi , ui ) � Xi � Ui
ei = Piui
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 26
Optimal Dispatch Problem
Uncertainty model• Weather• Occupancy• Prosumer consumption (PV & buildings)
minx,u
�E
�costi ·ei + �(xi � x̄) + �ei � (di + pi)�2
�
s.t. xi+1 = Aixi + Biui + Tiwi
(xi , ui) � Xi � Ui �wi � Wi
ei = Piui
Day-ahead plan for the thermal (x, e) trajectory of the building and electrical dispatch d
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 27
Prediction with Gaussian Processes (GP)Historical data
Estimates mean and uncertainty about the mean. Uncertainty of GP interpretation grows away from previous observations.
Train Gaussian Process
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 28
Optimal Dispatch Problem
Empirical distribution with samples extracted from Gaussian Process
minx,u
�E
�costi ·ei + �(xi � x̄) + �ei � (di + pi)�2
�
s.t. xi+1 = Aixi + Biui + Tiwi
(xi , ui) � Xi � Ui �wi � Wi
ei = Piui
Day-ahead plan for the thermal (x, e) trajectory of the building and electrical dispatch d
Solve using standard stochastic optimization
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 29
0 5 10 15 2018
20
22
24
26
28
Tem
pera
ture
0 5 10 15 2018
20
22
24
26
28
Tem
pera
ture
Dispatch Plan vs Prediction ConfidenceSunny Day with No Clouds Possibly Cloudy Day
0 5 10 15 20
Time (h)
0
10
20
30
40
Pow
er (k
W)
0 5 10 15 20
Time (h)
0
10
20
30
40
Pow
er (k
W)
Increased uncertainty results in higher energy expenditure and higher control authority
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 30
Real-Time Control – Every 5 minBattery SOC0≈ Dispatch error integral
Weather measurementand forecast wi
Building measurements& state estimate x0
minx,u
�
i
�SOCi �SOCREFi �2 +
�
j
�Tjxi ��
k
Tkxi�2
s.t. xi+1 = Aixi + Biui + Liwi
(xi , ui) � Xi � Ui
SOCi+1 = �SOCi + �(Piui � bi)
❶ Move battery to reference SOC ❷ All zones equal temperature
Enforce comfort constraints
Simple battery model
• Dispatch errors encoded in SOC → Restore SOC to nominal• No prosumer forecast → Distributed operations
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 31
TemperatureOccupancy
LightHumidity
Power measurement
Solar radiationOutside temp
aboratoire d’ utomatique emand esponse
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 32
LADR Experimental Configuration
❶ LADR has been virtually scaled-up in all experiments
PeakConsumption(kW)
LADR 45
Battery 720
Uncontrollable prosumers(PV+Buildings) 350
Scaled up from a true peak consumption of 7.8 kW
❷ Dispatch objective : Maximize flexibility
12% of loads are controllable
17-12 00:00 17-12 06:00 17-12 12:00 17-12 18:00 18-12 00:000
20
40
60
80
100
SOC
per
cent
age
[%]
Good Prediction Day – Battery Only
17-12 00:00 17-12 06:00 17-12 12:00 17-12 18:00 18-12 00:000
50
100
150
200
250
Pow
er[k
W]
ConsumptionDispatch
Dispatch plan (black)
Prosumer consumption (red)
SOC reference
Battery easily compensates for prediction errors (SOC)
17-12 00:00 17-12 06:00 17-12 12:00 17-12 18:00 18-12 00:00
20
22
24
2617-12 00:00 17-12 06:00 17-12 12:00 17-12 18:00 18-12 00:00
0
20
40
60
80
100
SOC
per
cent
age
[%]
17-12 00:00 17-12 06:00 17-12 12:00 17-12 18:00 18-12 00:000
50
100
150
200
250
Pow
er[k
W]
ConsumptionDispatch
Good Prediction Day – Battery & LADR
Building dispatch plan (black)
Building consumption (red)
SOC with battery (red)
Building zone temperatures
Building reduces battery requirements significantly
14-12 00:00 14-12 06:00 14-12 12:00 14-12 18:00 15-12 00:000
20
40
60
80
100
SOC
per
cent
age
[%]
14-12 00:00 14-12 06:00 14-12 12:00 14-12 18:00 15-12 00:000
100
200
300
Pow
er[k
W]
ConsumptionDispatch
Poor Prediction Day – Battery OnlyDispatch plan (black)
Prosumer consumption (red)
Battery cannot compensate for forecast errors (SOC)
14-12 00:00 14-12 06:00 14-12 12:00 14-12 18:00 15-12 00:00
20
22
24
2614-12 00:00 14-12 06:00 14-12 12:00 14-12 18:00 15-12 00:00
0
20
40
60
80
100
SOC
per
cent
age
[%]
14-12 00:00 14-12 06:00 14-12 12:00 14-12 18:00 15-12 00:000
100
200
300
Pow
er[k
W]
ConsumptionDispatch
Poor Prediction Day – Battery & LADR
Building dispatch plan (black)
Building consumption (red)
Building zone temperatures
Building reduces battery requirements significantly (SOC)
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 37
0 5 10 15 20 25 30 35Percentage controllable loads [%]
0
100
200
300
400
500
600
700
Req
uire
d Ba
ttery
Cap
acity
[kW
h]
00
20
40
60
80
100
120
Perc
enta
ge o
f Ins
talle
d Ba
tteryExperiments run at 12%
20% controllable buildings results in 80% reduction
in battery
Impact of Building Size – Preliminary Conclusion
Current battery
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 38
0 20 40 60 80 100 120Sample Period [min]
0
100
200
300
400
500
600
Req
uire
d Ba
ttery
Cap
acity
[kW
h]
00
20
40
60
80
100
Perc
enta
ge o
f Ins
talle
d Ba
ttery
Impact of Sample Rate – Preliminary Conclusion
Experiments run at 5min
Control of average energy over 60min
requires 20% larger battery
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 39
Conclusion
Powe
r
High
Energy
Low
High
Time
Fast
Slow
✓ ✓×
×✓Slow×
Spectral Split
Tracking signal
Conclusion: Hybrid storage schemes provide much greater flexibility to offer a wide variety of services at lower cost
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 40
Acknowledgements (i.e., they who did the work!)
Laboratoire d’Automatique Distributed Electrical Systems Laboratory
Luca Fabietti
Tomasz Gorecki
Fabrizio Sossan
Emil Namor
A dispatched-by-design architecture for distribution systems , Paolone-Jones | 06.02.2017 41
References[1] F. Sossan; E. Namor; R. Cherkaoui; M. Paolone, Achieving the Dispatchability of Distribution Feeders through Prosumers Data Driven Forecasting and Model Predictive Control of Electrochemical Storage, in IEEE Transactions on Sustainable Energy , 10.1109/TSTE.2016.2600103.
[2] E. Namor, F. Sossan, R. Cherkaoui and M. Paolone, Load Leveling and Dispatchability of a Medium Voltage Active Feeder through Battery Energy Storage Systems: Formulation of the Control Problem and Experimental Validation, in proceedings of ISGT Europe 2016, Ljubljana, Slovenija, October 9-12, 2016.
[3] F. Sossan and M. Paolone, Integration and Operation of Utility-Scale Battery Energy Storage Systems: the EPFL's Experience, in proceedings of Control of Transmission and Distribution Smart Grids, CTDSG 2016, Praha, October 11-13, 2016.