Flash flood prediction in large dams using neural...
1
Flash flood prediction in large dams using neural networks Flash flood prediction in large dams using neural networks JC 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 Introduction Training event (25-02-2003) Fi 11 h th f t A flow forecasting methodology is presented as a support tool for real time flood prediction in large dams. The practical and efficient use of hydrological real-time t i t t l i t f fl d di t The prediction models Th ANN dl h b t i d d lid t d f 12 fl d t ti td 90.0 100.0 110.0 120.0 130.0 140.0 150.0 m 3 /s) Caudal observado Predicción 3 Horas Modelo ANN C3C Training event (25-02-2003) 3 /s) 150 200 250 300 350 400 450 500 Q ANN (m 3 /s) C3C ANN - 1h training 100 150 200 250 300 Q ANN (m 3 /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 disasters prevention in catchments regulated with reservoirs. In this case, the optimal dam operation during flood scenarios should reduce the downstream risks and achieve a The ANN models have been trained and validated from 12 flood events estimated off-line (figure 4). A cross correlation analysis between precipitation data and inflows was previously performed for several historical events (figure 5) 00 10.0 20.0 30.0 40.0 50.0 60.0 70.0 80.0 Caudal (m Q (m 3 0 50 100 0 50 100 150 200 250 300 350 400 450 500 Q obs (m 3 /s) 0 50 0 50 100 150 200 250 300 Q obs (m 3 /s) 350 400 450 500 ) C3C ANN - 2h training 250 300 ) C3C ANN - 2h validation model for two recorded flood events used in the training and validation sets operation during flood scenarios should reduce the downstream risks and achieve a compromise between the structural security and the objectives of the water resources 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 BARTOLOME ALOSNO BURRILLA PEÑA AVERAGE 0.0 120 130 140 150 160 170 180 190 200 210 220 200.0 220.0 240.0 Caudal observado Predicción 3 Horas Modelo ANN C3C Validation event 17-12-2002 Time (hours) 0 50 100 150 200 250 300 350 0 100 200 300 400 500 Q obs (m 3 /s) Q ANN (m 3 /s) 0 50 100 150 200 0 50 100 150 200 250 300 Q obs (m 3 /s) Q ANN (m 3 /s) respectively. Additionally, the dispersion diagrams for the t t lit f t di th A dam operation during a flood event requires to take appropriate management strategies 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 event 0.10 0.20 0.30 0.40 0.50 Correlation coefficient PRESA SAN BARTOLOME ALOSNO BURRILLA PEÑA AVERAGE 0.10 0.20 0.30 0.40 0.50 Correlation coefficient 60.0 80.0 100.0 120.0 140.0 160.0 180.0 Caudal (m 3 /s) Modelo ANN C3C Q (m 3 /s) Qobs (m /s) 100 0 150.0 200.0 250.0 300.0 350.0 400.0 450.0 500.0 Q ANN (m 3 /s) C3C ANN - 3h training 50 0 100.0 150.0 200.0 250.0 300.0 Q ANN (m 3 /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 inherent stochastic character of peak rainfall intensities, their strong spatial and temporal i bilit d th hi hl li f id d i id th t the event. Additionally, an event based autocorrelation Figure 5. Cross correlation analysis for some events 0.00 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Lag time (hours) 0.00 0 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 50 110 120 130 140 150 Time (hours) 0.0 50.0 100.0 0.0 100.0 200.0 300.0 400.0 500.0 Q obs (m 3 /s) 0.0 50.0 0.0 50.0 100.0 150.0 200.0 250.0 300.0 Q obs (m 3 /s) sets are shown too. variability, and the highly nonlinear response of arid and semiarid catchments resulting from a high sensitivity to the soil moisture initial conditions and the dominant 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 system should include combined techniques of data pre-processing and completion coefficient near to 0.50 for a 5-hours lag time, i bl Analysis of QPF uncertainty in C3C model Above, it has been shown that C3C ANN model has good prediction capabilities. 0.20 0.30 0.40 0.50 0.60 0.70 Correlation coefficient 2003-02-25 2003-03-27 2003-04-22 2003-10-25 2003-11-22 2004-02-24 should include combined techniques of data pre processing and completion, assimilation of information and implementation of real time filters depending on the 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 Nevertheless, the proposed model has a high dependence on the QPF precipitation forecast. Usually, QPF have errors greater than 50 %, and therefore, it has been 0.00 0.10 1 2 3 4 5 6 Lag time (hours) C Average This work explores the capability of flood forecast algorithms based on artificial neural networks (ANN) techniques and their integration in a real time prediction training and validation process for some events Several ANN models architectures have been evaluated and compared All of analyzed what would be the effect of a systematic error of 50 % in QPF over the C3C forecast model performance. The results show a worsening of the NSE and RMSE indices especially in the case of QPF overestimation For this reason it has been tool developed that has been named PCTR, which is the Spanish acronym for “Real Time Flood Forecasting”. been evaluated and compared. All of them have a very simple architecture based on the conventional Three Layer indices, especially in the case of QPF overestimation. For this reason, it has been proposed a fourth ANN model (C3D) with greater precipitation time delays and thus eliminating this source of uncertainty. See figure 12. Methodology and case study based on the conventional Three Layer Feed Forward Perceptron, with a variable number 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 90 1.00 90 0 100.0 Validation set 0 90 1.00 45 0 50.0 The proposed forecasting methodology has been tested in the Meca River catchment (Huelva, Spain), l d b l S h d (f ) one single node in the output layer producing the next-hour flow value. structure (from F. Chang J. et al, 2007 ) Logistic activation 0 50 0.60 0.70 0.80 0.90 SE 50 0 60.0 70.0 80.0 90.0 E (m 3 /s) 0 50 0.60 0.70 0.80 0.90 SE 25 0 30.0 35.0 40.0 45.0 (m 3 /s) regulated byEl Sancho dam (figure 1). A hydrological data network of 5 telemetered rain ( 10 i) d 3 hi h ii l l A potential based transformation is applied to the original input and output variables for the ANN models function 0.10 0.20 0.30 0.40 0.50 NS 10.0 20.0 30.0 40.0 50.0 RMSE 0 10 0.20 0.30 0.40 0.50 NS 50 10.0 15.0 20.0 25.0 RMSE gauges (t=10 min) and 3 high precision water level sensors are operating in the catchment and reservoir with real-time data transmission to a variables for the ANN models. ANN Input variables preprocessing C3 and C3A ModelS C3B Model 0.00 1 h 2 h 3 h 4 h 5 h Time-ahead (hours) 0.0 0.00 0.10 1 h 2 h 3 h 4 h 5 h Time-ahead (hours) 0.0 5.0 reservoir, with real-time data transmission to a central database, making the data available at the dam control site. ANN Input variables preprocessing potential functions: P p Q q RMSE C3C P(ok) RMSE C3C P(-50%) RMSE C3C P(+50%) RMSE C3D NSE 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 C3D NSE 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, the same model with an assumed error of ±50 % and the proposed C3D model without QPF as input variable Conclusions The applied methodology includes an “on line” time series reconstruction of Figure 1. The Meca river catchment M P p M Q 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 of i f ll di ti i ti l h it i ti td hil t h h Conclusions historical output flows derived from the hydraulics of the gates (figure 2), while the inflows 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 when it is underestimated. To reduce such sensitivity, a new model (C3D) was proposed eliminating completely the predicted rainfalls in the input variables set Although The few-hours ahead inflows are predicted with an ANN model using as input variables a sequence of current and past average hourly inflows and rainfalls in the ff scheme is used following the strategy suggested 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. Although results are slightly poorer than in C3C model, the NSE index reveals a satisfactory performance 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 the immediate future quantitative precipitation forecast (QPF) from an outside model. see figure 8. The evaluated architectures of ANN models are shown in figure 9. Figure 9. Several ANN architectures evaluated: C3, C3A,C3B, C3C and C3D nets The robustness and simplicity of ANN schemes makes them particularly appropriate in real-time systems, as they can easily be integrated and programmed, handling Training and validation of ANN models The 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) and the Nash-Sutcliffe efficiency (NSE) statistical indices. The prediction horizon has been set to 3 hours although results show that it could be extended a few extra OUT IN Q t V Q are easily updated without changing the general conception and operation of the real-time decision making support tool. been set to 3 hours, although results show that it could be extended a few extra hours if the precipitation forecasts were reliable enough. Initially, it has been Figure 2 Three cases of weir flow and the corresponding discharge equations 5 . 1 1 u 1 OUT H L C Q 0.5 1 u 0 2 OUT 2 b H g 2 L G C Q 2 3 3 u OUT Z C g 2 L 3 2 Q ) Z ( V ) Z ( V V t t t l d i h ii d lid i Beven, K., Romanowicz R., Pappenberger, F., Young, P., and Werner, M., 2005. The Uncertainty Cascade in References Figure 2. Three cases of weir flow and the corresponding discharge equations In the above equation: is the mean inflow, is the storage variation and is the output reservoir IN Q V Q Statistical indices for model performance assessment: n evaluated inthe training and validation processes the statistical performance of 3 different ANN models : C3 C3A - C3B and C3C Between them 300 350 e = 1 mm e = 5 mm e = 10 mm Beven, K., Romanowicz R., Pappenberger, F., Young, P., and Werner, M., 2005. The Uncertainty Cascade in Flood 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 the River Arno Hydrological Sciences–Journal–des Sciences Hydrologiques 48(3) 381-398 the storage variation, and is the output reservoir discharge in the t interval. Figure 3, shows the high sensitivity of the estimation as a function of the OUT Q IN Q n Q Q RMSE i i i 1 2 ) ˆ ( n i i i n i i i Q Q Q Q NSE 1 2 1 2 ) ( ) ˆ ( 1 models : C3, C3A - C3B and C3C. Between them, the C3C model showed a better performance in the validation data set. See figure 10. 100 150 200 250 E(Q IN ) (m 3 /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 dl d l k f l d d level sensor precision e and the t. IN 0 50 0 10 20 30 40 50 60 t (minutos) Training set 0.95 1.00 25.0 30.0 Validation set 0.90 1.00 25.0 30.0 The Q OUT estimation sensitivity problem: Corani G. and Guariso G., 2005. Coupling Fuzzy Modeling and Neural Networks for River Flood Prediction. IEEE Transactions 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). Q Q Q E 54 90 54.95 55.00 55.05 55.10 55.15 tage (m) Z 2 Z + 2 Z 0.80 0.85 0.90 NSE 10.0 15.0 20.0 RMSE (m 3 /s) 0.60 0.70 0.80 NSE 10.0 15.0 20.0 RMSE (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. WaterResources Research 31 (10), 2517–2530. t e Z V e Z V t e Z V e Z V Q E Q Q Q E t t t t t t IN IN IN IN ) ( ) ( ) ( ) ( min max 54.65 54.70 54.75 54.80 54.85 54.90 Reservoir st Filtered water level (m) Measured water level (m) Z - 1 Z 1 e Z - 2 0.70 0.75 1 h 2 h 3 h Time-ahead (hours) 0.0 5.0 0.40 0.50 1 h 2 h 3 h Ti h d (h ) 0.0 5.0 Nash, J. E.y Sutcliffe, J. V., 1970. River flow forecasting through conceptual models. I. A discussion of principles. Journal of Hydrology, Vol. 10, p. 282-290. Zhang B Govindaraju R S 2003 Geomorphology based artificial neural networks (GANNs) for estimation of t t IN Fi 3 Fl i i i iii f diff ii () dh i i l Figure 10 NSE and RMSE statistical indices obtained for 3 ANN models Left: training data set 54.60 07/04/2002 00:00 07/04/2002 12:00 08/04/2002 00:00 08/04/2002 12:00 09/04/2002 00:00 09/04/2002 12:00 10/04/2002 00:00 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 of direct 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.
Transcript of Flash flood prediction in large dams using neural...
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) [email protected] ; (2) [email protected]
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.
