Post on 31-Jul-2020
Flash flood prediction in large dams using neural networksFlash flood prediction in large dams using neural networksJ C Múnera Estrada (1) and R García Bartual (2)J.C. Múnera Estrada (1) and R. García Bartual (2)
(1,2)Departamento de Ingeniería Hidráulica y Medio Ambiente, Universidad Politécnica de Valencia – Spain(1) j 1@d (2) i b@h(1) juamues1@doctor.upv.es ; (2) rgarciab@hma.upv.es
IntroductionTraining event (25-02-2003)
Fi 11 h th f t
A flow forecasting methodology is presented as a support tool for real time floodprediction in large dams. The practical and efficient use of hydrological real-time
t i t t l i t f fl d di tThe prediction models
Th ANN d l h b t i d d lid t d f 12 fl d t ti t d90.0
100.0
110.0
120.0
130.0
140.0
150.0
m3 /s
)
Caudal observado
Predicción 3 HorasModelo ANN C3C
Training event (25-02-2003)
3 /s)
150200250300350400450500
Q A
NN
(m3 /s
)
C3C ANN - 1h training
100
150
200
250
300
Q A
NN
(m3 /s
)
C3C ANN - 1h validation
Figure 11 shows the forecast response of the chosen C3C model for two recorded flood
measurements is necessary to operate early warning systems for flood disastersprevention in catchments regulated with reservoirs. In this case, the optimal damoperation during flood scenarios should reduce the downstream risks and achieve a
The ANN models have been trained and validated from 12 flood events estimatedoff-line (figure 4). A cross correlation analysis between precipitation data andinflows was previously performed for several historical events (figure 5) 0 0
10.0
20.0
30.0
40.0
50.0
60.0
70.0
80.0
Cau
dal (
mQ
(m3
050
100
0 50 100 150 200 250 300 350 400 450 500
Q obs (m3/s)
0
50
0 50 100 150 200 250 300
Q obs (m3/s)
350400450500
)
C3C ANN - 2h training250
300
)
C3C ANN - 2h validationmodel for two recorded flood events used in the training and validation sets
operation during flood scenarios should reduce the downstream risks and achieve acompromise between the structural security and the objectives of the waterresources system management.
inflows was previously performed for several historical events (figure 5).
Optimal time lags were f d b i h f
2002-12-17 event
0 50
0.60
0.70
(k)
2002-04-08 event
0 50
0.60
0.70
(k)
PRESA SAN BARTOLOMEALOSNO BURRILLAPEÑA AVERAGE
0.0120 130 140 150 160 170 180 190 200 210 220
Patrón (horas)
200.0
220.0
240.0Caudal observado
Predicción 3 HorasModelo ANN C3C
Validation event 17-12-2002Time (hours)
050
100150200250300350
0 100 200 300 400 500
Q obs (m3/s)
Q A
NN
(m3 /s
)
0
50
100
150
200
0 50 100 150 200 250 300
Q obs (m3/s)
Q A
NN
(m3 /s
)
respectively. Additionally, the dispersion diagrams for the t t lit f t d i th
y gA dam operation during a flood event requires to take appropriate managementstrategies depending on the flood magnitude and the initial freeboard at the
found to be in the range of 2 to 6 hours, depending on the event0.10
0.20
0.30
0.40
0.50
Cor
rela
tion
coef
ficie
nt
PRESA SAN BARTOLOMEALOSNO BURRILLAPEÑA AVERAGE
0.10
0.20
0.30
0.40
0.50
Cor
rela
tion
coef
ficie
nt
60.0
80.0
100.0
120.0
140.0
160.0
180.0
Cau
dal (
m3 /s
)
Modelo ANN C3C
Q(m
3 /s)
Q obs (m /s)
100 0150.0200.0250.0300.0350.0400.0450.0500.0
Q A
NN
(m3 /s
) C3C ANN - 3h training
50 0
100.0
150.0
200.0
250.0
300.0
Q A
NN
(m3 /s
)
C3C ANN - 2h validation
totality of patrons used in the training and validation data sets are shown too
reservoir. The most important flow prediction difficulties arise from the inherentstochastic character of peak rainfall intensities, their strong spatial and temporal
i bilit d th hi hl li f id d i id t h t
the event. Additionally, an event based autocorrelation
Figure 5. Cross correlation analysis for some events
0.000 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Lag time (hours)
0.000 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Lag time (hours)
Figure 11. Left: Two examples of estimated vs forecast flood events
0.0
20.0
40.0
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150
Patrón (horas)Time (hours)
0.050.0
100.0
0.0 100.0 200.0 300.0 400.0 500.0
Q obs (m3/s)
0.0
50.0
0.0 50.0 100.0 150.0 200.0 250.0 300.0
Q obs (m3/s)
sets are shown too. variability, and the highly nonlinear response of arid and semiarid catchmentsresulting from a high sensitivity to the soil moisture initial conditions and thedominant flow mechanisms
based autocorrelation analysis (figure 6) shows an average correlation
A QPF C3C 0.80
0.90
1.00
t (
k) 2002-04-08
2002-12-17
Event Date
from training set (up) and validation set (down) Right: Comparison of estimated and forecast flows for 1h to 3h time-ahead in both series.
dominant flow mechanisms.The efficient integration of a flow forecast model in a real-time prediction systemshould include combined techniques of data pre-processing and completion
gcoefficient near to 0.50 for a 5-hours lag time,
i bl
Analysis of QPF uncertainty in C3C modelAbove, it has been shown that C3C ANN model has good prediction capabilities.0.20
0.30
0.40
0.50
0.60
0.70
Cor
rela
tion
coef
ficie
nt 2003-02-25
2003-03-27
2003-04-22
2003-10-25
2003-11-22
2004-02-24should include combined techniques of data pre processing and completion,assimilation of information and implementation of real time filters depending onthe system characteristics.
suggesting a reasonable prediction horizon.Figure 4. Some recorded flood
events used in the ANN model training and validation process
Figure 6. Autocorrelation analysis for some events
, g p pNevertheless, the proposed model has a high dependence on the QPF precipitationforecast. Usually, QPF have errors greater than 50 %, and therefore, it has been
0.00
0.10
1 2 3 4 5 6Lag time (hours)
C
Average
This work explores the capability of flood forecast algorithms based on artificialneural networks (ANN) techniques and their integration in a real time prediction
training and validation process for some events
Several ANN models architectures havebeen evaluated and compared All of
analyzed what would be the effect of a systematic error of 50 % in QPF over the C3Cforecast model performance. The results show a worsening of the NSE and RMSEindices especially in the case of QPF overestimation For this reason it has beentool developed that has been named PCTR, which is the Spanish acronym for “Real
Time Flood Forecasting”.been evaluated and compared. All ofthem have a very simple architecturebased on the conventional Three Layer
indices, especially in the case of QPF overestimation. For this reason, it has beenproposed a fourth ANN model (C3D) with greater precipitation time delays and thuseliminating this source of uncertainty. See figure 12.
Methodology and case studybased on the conventional Three LayerFeed Forward Perceptron, with a variablenumber of nodes in the hidden layer and Figure 8. Serial propagation neural networks
structure (from F Chang J et al 2007 )
eliminating this source of uncertainty. See figure 12.Training set
0 901.00
90 0100.0
Validation set
0 901.00
45 050.0
The proposed forecasting methodology has beentested in the Meca River catchment (Huelva, Spain),
l d b l S h d (f )
one single node in the output layerproducing the next-hour flow value.
structure (from F. Chang J. et al, 2007 )
Logistic activation 0 50
0.600.700.800.90
SE 50 060.070.080.090.0
E (m
3 /s)
0 500.600.700.800.90
SE 25 030.035.040.045.0
(m3 /s
)
regulated by El Sancho dam (figure 1).A hydrological data network of 5 telemetered rain
( 10 i ) d 3 hi h i i l l
A potential based transformation isapplied to the original input and outputvariables for the ANN models
function
0.100.200.300.400.50
NS
10.020.030.040.050.0
RM
SE
0 100.200.300.400.50
NS
5 010.015.020.025.0
RM
SE
gauges (t=10 min) and 3 high precision water levelsensors are operating in the catchment andreservoir with real-time data transmission to a
variables for the ANN models.
ANN Input variables preprocessingC3 and C3A
ModelS C3B Model
0.000 0
1 h 2 h 3 h 4 h 5 h
Time-ahead (hours)
0.00 0
0.000.10
1 h 2 h 3 h 4 h 5 hTime-ahead (hours)
0.05.0
reservoir, with real-time data transmission to acentral database, making the data available at thedam control site.
ANN Input variables preprocessing potential functions:
Pp
RMSE C3C P(ok) RMSE C3C P(-50%) RMSE C3C P(+50%) RMSE C3DNSE C3C P(ok) NSE C3C P(-50%) NSE C3C P(+50%) NSE C3D
RMSE C3C P(ok) RMSE C3C P(-50%) RMSE C3C P(+50%) RMSE C3DNSE C3C P(ok) NSE C3C P(-50%) NSE C3C P(+50%) NSE C3D
Figure 12. NSE and RMSE statistical indices obtained for C3C model with an hypothetical perfect QPF, thesame model with an assumed error of ±50 % and the proposed C3D model without QPF as input variable
ConclusionsThe applied methodology includes an “on line” time series reconstruction of
Figure 1. The Meca river catchment
MP
p
MQ
q same model with an assumed error of ±50 % and the proposed C3D model without QPF as input variable.Left: training data set. Right: validation data set.
Quality of predictions has been found to be strongly affected by reliability ofi f ll di ti i ti l h it i ti t d hil t h h
Conclusionshistorical output flows derived from the hydraulics of the gates (figure 2), while theinflows estimation is made by means of mass balance equation in the reservoir.
For the following time steps, a serial-propagated neural networks structure
rainfall predictions, in particular when it is overestimated, while not so much whenit is underestimated. To reduce such sensitivity, a new model (C3D) was proposedeliminating completely the predicted rainfalls in the input variables set Although
The few-hours ahead inflows are predicted with an ANN model using as inputvariables a sequence of current and past average hourly inflows and rainfalls in the
ff
scheme is used following the strategysuggested by F. Chang J. et al (2007),see figure 8 The evaluated architectures Fi 9 S l ANN hi l d
C3C Model C3D Model
eliminating completely the predicted rainfalls in the input variables set. Althoughresults are slightly poorer than in C3C model, the NSE index reveals a satisfactoryperformance in the validation set (near 0.80 for 2 hours and 0.60 for 3 hours).
catchment, each one with different time delays. Moreover, it is included theimmediate future quantitative precipitation forecast (QPF) from an outside model.
see figure 8. The evaluated architecturesof ANN models are shown in figure 9.
Figure 9. Several ANN architectures evaluated:C3, C3A,C3B, C3C and C3D nets
pThe robustness and simplicity of ANN schemes makes them particularly appropriatein real-time systems, as they can easily be integrated and programmed, handlingTraining and validation of ANN modelsThe mass balance
equation:well the presence of possible errors and uncertainties in data.On the other hand, this models are computationally very efficient, and over all, they
The ANN models have been compared using the Root Mean Square Error (RMSE) andthe Nash-Sutcliffe efficiency (NSE) statistical indices. The prediction horizon hasbeen set to 3 hours although results show that it could be extended a few extra
OUTIN QtVQ
are easily updated without changing the general conception and operation of thereal-time decision making support tool.
been set to 3 hours, although results show that it could be extended a few extrahours if the precipitation forecasts were reliable enough. Initially, it has been
Figure 2 Three cases of weir flow and the corresponding discharge equations
5.11u1OUT HLCQ
0.5
1u02OUT 2bHg2LGCQ
2
33uOUT ZCg2L
32Q )Z(V)Z(VV ttt
l d i h i i d lid i
Beven, K., Romanowicz R., Pappenberger, F., Young, P., and Werner, M., 2005. The Uncertainty Cascade inReferences
Figure 2. Three cases of weir flow and the corresponding discharge equations
In the above equation: is the mean inflow, isthe storage variation and is the output reservoir
INQ VQ
Statistical indices for model performance assessment:
n
evaluated in the training and validation processesthe statistical performance of 3 different ANNmodels : C3 C3A - C3B and C3C Between them300
350e = 1 mm e = 5 mm e = 10 mm
Beven, K., Romanowicz R., Pappenberger, F., Young, P., and Werner, M., 2005. The Uncertainty Cascade inFlood Forecasting. International conference on innovation FloodRelief. Bergen-Tromsø, Norway. 9p.Campolo, M., Soldati, A., Andreussi, P., 2003. Artificial neural network approach to flood forecasting in theRiver Arno Hydrological Sciences–Journal–des Sciences Hydrologiques 48(3) 381-398
the storage variation, and is the output reservoirdischarge in the t interval. Figure 3, shows the highsensitivity of the estimation as a function of the
OUTQ
INQn
QQRMSE i
ii
1
2)ˆ(
n
i ii
n
i ii
QQNSE
12
12
)(
)ˆ(1
models : C3, C3A - C3B and C3C. Between them,the C3C model showed a better performance inthe validation data set. See figure 10.100
150
200
250
E(Q
IN) (
m3 /s
)
River Arno. Hydrological Sciences Journal des Sciences Hydrologiques, 48(3). 381 398.Chang F. J., Chiang, Y. M., Chang, L. C., 2007. Multi-step-ahead neural networks for flood forecasting.Hydrological Sciences–Journal–des Sciences Hydrologiques, 52(1). 114-130.C G d G G 200 C l d l d l k f l d d
ylevel sensor precision e and the t.
INQ g0
50
0 10 20 30 40 50 60
t (minutos)
Training set
0.95
1.00
25.0
30.0Validation set
0.90
1.00
25.0
30.0
The QOUT estimation sensitivity problem: Corani G. and Guariso G., 2005. Coupling Fuzzy Modeling and Neural Networks for River Flood Prediction. IEEETransactions on systems, man, and cybernetics—Part C: Applications and reviews, 35 (3). 382-390.García Bartual, R., Múnera, J. C. 2007. A Decision Support Tool for Flash Flood Control in Large Dams. (Poster). QQQE 54 90
54.9555.0055.05
55.1055.15
tage
(m)
Z2
Z+2
Z0.80
0.85
0.90
NSE
10.0
15.0
20.0
RM
SE (m
3 /s)
0.60
0.70
0.80
NSE
10.0
15.0
20.0
RM
SE (m
3 /s)
EGU General Assembly 2007. EGU2007-A-11012. Viena.Hsu, K., Gupta, H., Sorooshain, S., 1995. Artificial neural network modeling of the rainfall-runoff process.Water Resources Research 31 (10), 2517–2530.
t
eZVeZVt
eZVeZVQE
QQQE
ttttttIN
INININ
)()()()(
minmax
54.65
54.70
54.7554.80
54.8554.90
Res
ervo
ir st
Filtered water level (m)
Measured water level (m)Z-1
Z1e
Z-2
0.70
0.75
1 h 2 h 3 hTime-ahead (hours)
0.0
5.0
0.40
0.50
1 h 2 h 3 h
Ti h d (h )
0.0
5.0
ate esou ces esea c 3 ( 0), 5 530Nash, J. E.y Sutcliffe, J. V., 1970. River flow forecasting through conceptual models. I. A discussion ofprinciples. Journal of Hydrology, Vol. 10, p. 282-290.Zhang B Govindaraju R S 2003 Geomorphology based artificial neural networks (GANNs) for estimation of
ttIN
Fi 3 Fl i i i i i i f diff i i ( ) d h i i l Figure 10 NSE and RMSE statistical indices obtained for 3 ANN models Left: training data set
54.60
07/0
4/20
0200
:00
07/0
4/20
0212
:00
08/0
4/20
0200
:00
08/0
4/20
0212
:00
09/0
4/20
0200
:00
09/0
4/20
0212
:00
10/0
4/20
0200
:00
Date and Time
Measured water level (m)1 Time-ahead (hours)
C3 model (RMSE) C3C model (RMSE) C3A - C3B model (RMSE)
C3 model (NSE) C3A - C3B model (NSE) C3C model (NSE)
Time-ahead (hours)
C3 model (RMSE) C3A - C3B model (RMSE) C3C model (RMSE)
C3 model (NSE) C3A - C3B model (NSE) C3C model (NSE)
European geosciences union – EGU 2009 – Vienna, April 20-24. Poster Session HS10 5 11456
Zhang, B., Govindaraju R. S., 2003. Geomorphology-based artificial neural networks (GANNs) for estimation ofdirect runoff over watersheds. Journal of Hydrology 273. 18–34.
Figure 3. Flow input estimation sensitivity for different sensor precision (e) and the t compute time interval. Figure 10. NSE and RMSE statistical indices obtained for 3 ANN models. Left: training data set.Right: validation data set.