Urban flood forecast based on raingauge networks · 12 nd International Conference on Urban Drainage, Porto Alegre/Brazil, 10-16 September 2011 2 Urban flood forecast based on raingauge

12nd

International Conference on Urban Drainage, Porto Alegre/Brazil, 10-16 September 2011

Simões et al. 1

Urban flood forecast based on raingauge networks

N. Simões

1,2,*, L. Wang

1, S. Ochoa

1, J. P. Leitão

3, R. Pina

4, A. Sá Marques

2, Č. Maksimović

1,

R.F. Carvalho2, L. David

3

1 Department of Civil and Environmental Engineering, Imperial College London, Skempton building, South Kensington Campus, London, SW7 2AZ, United Kingdom

2 Department of Civil Engineering, University of Coimbra, Rua Luís Reis Santos, 3030-788 Coimbra, Portugal.

3 Laboratório Nacional de Engenharia Civil (LNEC), Av. do Brasil 101, 1700-066 Lisboa, Portugal.

4 AC, Águas de Coimbra, E.E.M., Rua da Alegria n8 111, 3000-018 Coimbra, Portugal *Corresponding author ([email protected], [email protected])

ABSTRACT

For reliable prediction of urban pluvial flooding it is essential to have reliable spatial and

temporal rainfall prediction at an appropriate scale. Radar data are considered to be the most

reliable. However, many urban catchments do not have access to radar data. This paper

presents a new methodology for rainfall forecasting based on a network of raingauges. The

methodology predicts rainfall at each raingauge location based on the Support Vector

Machine (SVM) technique improved with Singular Spectrum Analysis (SSA). The prediction

of the spatial distribution of rainfall is based on interpolation techniques. The forecasted

rainfall fields are then used as inputs for simulation of drainage systems to obtain a short-term

flood prediction. The proposed methodology is tested using real data from a case study in

Coimbra, Portugal and the results obtained showed that it is possible to predict the water level

30 minutes in advance when using this methodology.

KEYWORDS Pluvial flooding, flood modelling, rainfall forecasting, raingauges network

INTRODUCTION

Surface water flooding occurs due to extreme rainfall and the inability of the sewer system to

cope with all runoff. As a consequence, a considerable volume of water is carried out over the

surface through preferential flow paths and can eventually accumulate in natural or man-made

depressions/ponds. This can cause minor material losses but also major incidents with obvious

consequences in economic activities, human health (especially in combined systems) and

people’s lives (sometimes with fatal consequences). Due to climate change and increase of

urbanisation, urban flooding phenomena are happening and being reported more frequently.

One of the most critical aspects of flood forecast is the lead time, which corresponds to the

time period between the acquisition of data and the flood prediction results obtained from

hydraulic simulations. For this reason, data transmission, correction of anomalies and flood

hydraulic simulations need to be fast, reliable and as accurate as possible, in order to obtain

reliable and timely estimations of flood magnitude and extent. This would enable timely

triggering of alarms, warnings and/or real time control of urban drainage systems (Cembrano

et al, 2004, Schutze et al, 2004), in order to minimise the negative effects of these events

12nd


2 Urban flood forecast based on raingauge networks

Urban surface water flooding occurs at a small scale and is affected by the local topography,

the drainage infrastructure and the built urban environment. The events that cause this type of

flooding are characterised by rapid onset and high-intensity precipitation. For this reason, a

new reliable approach for urban surface water flood modelling and prediction is required.

This new approach has to include: i) short-term rainfall analysis and prediction, and ii) short-

term flood prediction. These two modules must be coupled together, so that surface water

flooding can be accurately and timely modelled (for example Maksimovic et al. (2009)) and

predicted (for example, Leitão et al. (2010)).

In urban catchments the time interval between the measurement (onset) of rainfall data and

the occurrence of flood peaks is in the majority of the cases less than 30-60 minutes. In order

to improve the efficiency and the predictability of observations and forecasting, radar data is

becoming widely used; however, some areas in the world are unable to afford or access this

type of data. In these cases it is possible to rely on the network of raingauges which can

provide some sort of spatially distributed rainfall field and thus the associated forecasting can

be based merely on using rain gauge data which need to be developed and tested. The major

disadvantage of raingauge measurements is the lack of detailed spatial information of rainfall.

It is therefore crucial to have at least a few raingauges at appropriate distance to the catchment

in analysis and to reproduce the spatial variability of real rain fields based on individual rain

gauge sites.

Interpolation techniques are a possible solution to reproducing the spatial and temporal

variability of rain fields. Some breakthroughs have been made to improve the conventionally

widely-used interpolation methods, such as Inverse Distance Weight, Kriging and Cubic

splines. Bárdossy & Pegram (2010), for example, used the mathematical technique Copulas

and the Circulation Patterns to help characterising the spatial structure of rainfall and link

coarser weather information to refine the Kriging interpolation method. Vischel et al. (2010)

developed a simple dynamic interpolation technique that incorporates the kinematics of the

rainfall systems; the results demonstrated the ability to produce more feasible spatial

distribution of rainfall.

For very short lead times, quantitative precipitation forecasting can be achieved using

extrapolation of consecutive images (e.g. using Support Vector Machines (SVM) (Gupta et al., 2009), Artificial Neural Networks (ANN) (Hung et al., 2009), Auto-Regressive Moving

Average (ARMA) models (Burlando et al., 1993)). Behind the ANN and SVM methods

stands the concept of the learning model. Several authors obtained good results with the

application of SVM: Dibike et al. (2001) demonstrated the capability of SVM in hydrological

prediction for modelling the rainfall--runoff process and found that the SVM produced better

prediction of runoff on test data when compared to the ANN model; Gupta et al. (2009)

applied SVM to forecast rainfall with a lead time from 15- to 30-min by integrating and

analysing the rainfall data of three consecutive years in Mumbay.

This paper presents a new methodology for rainfall forecasting based on raingauge networks.

The proposed approach comprised two steps: (1) prediction of rainfall in each raingauge

based on a SVM algorithm; and (2) prediction of the spatial distribution of rainfall based on

the application of interpolation techniques to the forecasted values obtained in step (1). The

forecasted rainfall fields are then used as inputs for simulate the hydraulic conditions of the

urban drainage systems based on the dual-drainage model (Maksimović et al., 2009), in order

to obtain a short-term flood prediction. The proposed methodology is tested using real data

from a case study in Coimbra, Portugal.

12nd


Simões et al. 3

METHODOLOGY

A methodology for urban flood forecast based on raingauge networks is presented (see Figure

1). The rainfall forecast is done using a SSA-SVM methodology (Simões et al., 2011). Using

an interpolation technique the rain fields are generated (Figure 2) and then used in the 1D/1D

hydraulic model.

Figure 1: Flowchart of the developed

methodology

Figure 2: Rainfall forecast and generation of

rainfields.

Case Study

The Portuguese city of Coimbra is a medium size city that experienced several urban floods in

the recent years. The catchment studied herein has a total area of approximately 1.5km2 and

discharges in the Coselhas brook. The area of the catchment where the main flood problems

occur is highly urbanised (“Zona Central”) with approximately 0.9 km2 area (Figure 3).

Figure 3: Digital terrain model of “Zona Central” Catchment

The sewer system is 34.8 km long, of which 29 km are combined sewers and only 1.2 km are

for storm water only. The time of concentration of the catchment (that contributes to the flood

area) is estimated to be 45 minutes.

Monitoring campaign

12nd

International Conference on Urban Drainage, Porto Alegre/Brazil, 10

4

Full-scale test on a experimental site in Coi

the methodology presented in this paper. Three tipping bucket raingauges were installed, as

well as three level gauges in the sewer system and one levelgauge in the surface (Figure 4).

Figure 4: Case study area and location of raingauges and level gauges

Rainfall Prediction

Support Vector Machines were developed with the objective to solve pattern recognition and

classification problems. This

regression estimation problems and have also been successfully applied to solve forecasting

problems in many fields.

The SVMlight

software was used in this work (Joachims 1999).

SVMlight

is used in this work to carry out the time series prediction. Implemented based upon

Vapnik (1995) and Joachims (1999, 2002), the SVM

from Department of Computer Science in Cornell University.

Singular Spectrum Analysis is a non

(Sivapragasam et al., 2001). Its usefulness has been proven in the analysis of climatic,

meteorological and geophysical time series (Alonso et al., 2005). It separates a data series i

two data series: one smoothed and its residuals (data = smoothed series + residuals).

Singular Spectrum Analysis is applied to all data series and two data sets are generated: the

smoothed series and the residual. Each of the data series will generate t

the prediction is done separately for each series. At the end, the smoothed and residuals

predictions are added and the rainfall predictions are obtained.

Raingauges

Levelgauges

International Conference on Urban Drainage, Porto Alegre/Brazil, 10-


scale test on a experimental site in Coimbra (Zona Central) is being carried out to assess



: Case study area and location of raingauges and level gauges


This technique has been further extended to solve nonlinear


software was used in this work (Joachims 1999). The regression module of the


Vapnik (1995) and Joachims (1999, 2002), the SVMlight

was developed by Thorsten Joachims

from Department of Computer Science in Cornell University.

alysis is a non-parametric technique used in the analysis of time series


meteorological and geophysical time series (Alonso et al., 2005). It separates a data series i



smoothed series and the residual. Each of the data series will generate two SVM models and


predictions are added and the rainfall predictions are obtained.

P. Republica

Avenida

Mercado

-16 September 2011


mbra (Zona Central) is being carried out to assess



: Case study area and location of raingauges and level gauges


en further extended to solve nonlinear


The regression module of the


was developed by Thorsten Joachims

parametric technique used in the analysis of time series


meteorological and geophysical time series (Alonso et al., 2005). It separates a data series in



wo SVM models and


12nd


Simões et al. 5

Generation of Rainfall fields

Using different interpolation techniques, it is possible to estimate rainfall spatial variability

from rain gauge data and to obtain a “virtual radar image” and then, apply some extrapolation

and more advanced techniques to make short-term rainfall forecast, such as inverse distance

interpolation. The Inverse Distance Interpolation method weights every grid point according

to its distance from the sample point.

( ) 0

0

( )

( )

n

k kkn

kk

w x uu x

w x=

=

=∑

∑

where is the value of the sample point , and the weight associated depends of the distance and

is defined as:

( )( )

1

,k

k

w xd x x

=

In this method the extremes are typically located at the data points which results in poor shape

properties.

Hydraulic model

The Urban Water Research Group (UWRG) from Imperial College London developed a tool

based on the AOFD methodology comprising several GIS-based routines that automatically

analyses, quantifies and generates 1D overland flow networks (ponds and flow paths). The

tool analyses several GIS layers, such as Digital Elevation (Maksimovic et al, 2009).

The 1D/1D model was then created by employing the storage nodes and overland flow paths

delineated using the AOFD methodology (Maksimovic et al. 2009). A LiDAR (Light

Detection and Ranging) DEM with cell size 1x1m and vertical resolution of approximately

0.15m was used in the delineation. The cross-sections of the overland flow paths were

confined to open rectangular or open trapezoidal channels. The 1D/1D model was then set up

by coupling the sewer network and the 1D overland flow network.

RESULTS

The monitoring campaign recorded a high rainfall event in 08-10-2010. Using the data of this

event, the complete methodology, as proposed in this paper, for urban flood forecast based on

a network of raingauges was tested. Results showed that, with this methodology it was

possible to predict the water level in sewers 30 minutes in advance. Figure 5 presents the

rainfall recorded in each of the 3 raingauges used.

Figure 5: Rainfall recorded in 3 raingauges on the 8/10/2010 event.

0

10

20

30

40

50

60

70

15:50 16:19 16:48 17:16 17:45 18:14 18:43

Ra

infa

ll in

ten

sity

[mm

/hh

]

time [min]

R1

R2

R3

12nd



Figure 6 shows the interpolation of three consecutive rainfields based on the forecast result of

the 3 raingauges (forecast start time: 17h10min) and the resulting spatial interpolation.

17h25m

17h30min

17h35min

Figure 6: Three consecutive rainfields based on the forecast result of the 3 raingauges

(forecast start time: 17h10min) and the resulting spatial interpolation

Figure 7 shows the average of the rainfall recorded in the 3 raingauges and the forecasted

rainfall (average of 3 raingauges) with different forecast start time (fst).

Figures 8, 9 and 10 show the simulated water level in 3 different locations (P. República,

Avenida and Mercado) using the rainfall input (rainfields) of different forecasts. Figure 8 also

compares the simulation results obtained with the observed data in the sewer.

Figure 7: Rainfall (average of raingauges)

Figure 8: water level in Praça da Republica

Figure 9: Water level in Mercado

Figure 10: Water level in Avenida

0

5

10

15

20

25

30

35

40

45

50

16

:19

16

:33

16

:48

17

:02

17

:16

17

:31

17

:45

18

:00

Ra

infa

ll I

nte

nsi

ty [

mm

/hh

]

Time

Obs Rainfall

fst: 17:20

fst: 17:15

fst: 17:10

fst: 17:05

fst: 17:00

0.12

0.17

0.22

0.27

0.32

0.37

0.42

0.47

16

:15

16

:43

17

:12

17

:41

18

:10

Wa

ter

De

pth

[m

]

Time

Obs Rainfall

fst: 17:20

fst: 17:15

fst: 17:10

fst: 17:05

fst: 17:00

observed

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

0.55

0.6

16

:15

16

:43

17

:12

17

:41

18

:10

Wa

ter

De

pth

[m

]

Time

Obs Rainfall

fst: 17:20

fst: 17:15

fst: 17:10

fst: 17:05

fst: 17:00

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

16

:15

16

:43

17

:12

17

:41

18

:10

Wa

ter

De

pth

[m

]

Time

Obs Rainfall

fst: 17:20

fst: 17:15

fst: 17:10

fst: 17:05

fst: 17:00

12nd


Simões et al. 7

The difference between the peak value in the pipes is less than 3 cm; however the

hydrographs show a smoother curve for the simulations with forecasted data. This happens

because the rainfall prediction is also smoother than the observed data. Nevertheless, in terms

of flood prediction, the most important variable is water level.

DISCUSSION AND CONCLUSIONS

A methodology for urban flood forecast based on raingauge networks is presented in this

paper. The rainfall forecast is conducted using a SSA-SVM methodology, spatial interpolation

for generation of rainfields and then used it in the 1D1D hydraulic model (Figure 2).

The methodology was tested against observed data and it was possible to predict the water

level with 30 minutes in advance. This result showed that it is possible to have a flood alert

system based in a network of raingauges. Future tests with a denser raingauge networks need

to be done in order to increase the flood forecasting lead time.

ACKNOWLEDGEMENT

The research was conducted as part of the Flood Risk Management Research Consortium

(FRMRC2, SWP3). The authors would like to thank MWH Soft the provision of the software.

Nuno Simões acknowledges the financial support from the Fundação para a Ciência e

Tecnologia - Ministério para a Ciência, Tecnologia e Ensino Superior, Portugal

[SFRH/BD/37797/2007]. Li-Pen Wang acknowledges the full financial support of the

Ministry of Education Taiwan for his postgraduate research programme.

REFERENCES

Alonso, F.J., Del Castillo, J.M, Pintado, P., (2005), Application of singular spectrum analysis

to the smoothing of raw kinematic signals. J. Biomech. 38, 1085-1092

Bárdossy, A. and Pegram, G. (2010) Interpolation and simulation of daily rainfall observations based on Copulas and Circulation Patterns, Proceedings of ‘‘International

Workshop on Advances in Statistical Hydrology.’’ Taomina, Italy.

Burlando, P., Rosso, R., Cadavid L. G. and Salas, J. D.(1993) Forecasting of short-term rainfall using ARMA models. Journal of Hydrology, 144 (1-4), 193-211.

Cembrano, G., Quevedo, J., Salamero, M., Puig, V., Figueras, J., Marti, J.. Optimal control of

urban drainage systems. A case study. Control Engineering Practice, vol. 12, no. 1, pp. 1–9,

2004

Dibike, Y. B., Velickov, S., Solomatine D., and Abbott, M. B. (2001) Model induction with support vector machines: introduction and applications. Journal of Computing in Civil Eng,

15(3), 208–216.

Gupta, K. and Nikam V., Rainfall forecast for extreme monsoon rainfall conditions for urban

area, Proceedings of ‘‘8th International Workshop On Precipitation In Urban Areas.’’ St.

Moritz, Switzerland, 2009.

12nd



Hung, N. Q., Babel, M. S., Weesakul S., and Tripathi, N. K., An artificial neural network model for rainfall forecasting in Bangkok, Thailand. Hydrology and Earth System Science,

13, 1413–1425, 2009.

Joachims T. (1999). Making large-Scale SVM Learning Practical. Advances in Kernel

Methods. In: B. Schölkopf, C. Burges and A. Smola (eds.), Support Vector Learning, MIT-

Press, Chapter 11, pp. 41-56.

Joachims T. (2002). Learning to Classify Text Using Support Vector Machines. Kluwer

Academic Publishers, NL.

Leitao, J. P. , Simoes, N.E., Maksimovic, C, Ferreira, F., Prodanovic, D ,Matos, J.S. and Sa

Marques, A. Real-time forecasting urban drainage models: full or simplified networks?,

Water Science & Technology—WST, 62.9, 2010

Maksimović, Č., Prodanović, D., Boonya-aroonnet, S., Leitão, J. P., Djordjević, S., and Allitt,

R. (2009). Overland flow and pathway analysis for modelling of urban pluvial flooding.

Journal of Hydraulic Research, 47(4):512-523.

Simoes, N., Wang, L.,Ochoa, S. Leitão, J.P., Pina,R., Onof, C. Sá Marques, A., Maksimović,

C., Carvalho, R., David, L.,(2011) A coupled SSA-SVM technique for stochastic short-term

rainfall forecasting, 12nd

International Conference on Urban Drainage, Porto Alegre/Brazil,

10-16 September (submitted)

Sivapragasam, C., Liong, S. and Pasha, M., Rainfall and runoff forecasting with SSA–SVM

approach, Journal of Hydroinformatics, 03.3, 2001

Schütze M., Campisano A., Colas H., Schilling W. and Vanrolleghem P. : Real-time control of

urban wastewater systems - where do we stand today?, Journal of Hydrology, 299, 335-348

(2004)

Vapnik V. N. (1995). The nature of statistical learning theory. Springer, New York.

Vischel, T., Quantin, G., Level, T., Gosset, M. and Cazenave, F. (2010) Generation of high resolution rain fields in West Africa: 1. Evaluation of a dynamical interpolation technique, Proceedings of ‘‘10th International Precipitation Conference.’’ Coimbra, Portugal, 2010.

Documents

Urban flood forecast based on raingauge networks · 12 nd International Conference on Urban Drainage, Porto Alegre/Brazil, 10-16 September 2011 2 Urban flood forecast based on raingauge