Short term river flood forecasting with neural networks

This paper reports results obtained using artificial neural networks (ANN) models for shortterm river flow forecasting under heavy rain storms, in the upper Serpis river basin (460 km2), with the outlet in Beniarrés reservoir (29 hm3 ). The system is monitored by 6 raingauges, providing 5-min rainfall intensities, while reservoir inflows are derived from depth measurements in the reservoir every half hour and realtime data from controlled discharges in the spillway. In order to produce 1, 2 and 3 hours forecasts, the model makes use of the distributed rainfall information, together with observed discharges in the preceding hours. Several ANN topologies have been tested and compared, including linear and non linear schemes, being in all cases three-layer feedforward networks. Best results are obtained with different architectures for each forecasting horizon, basically due to the decreasing dependence of future inflows with respect to preceding values of the series as the time horizon is increased, while rainfall information increases its importance as a predictor. The ANN architectures finally proposed are achieved through pruning algorithms. Training is performed using the quasi-Newton approach. A specific software was also develop to help in the real-time management of the dam during floods, which incorporates all the relevant information about the dam elements (Tainter-gates, depth-volume relationships,...), and is designed for real-time operation, accepting as inputs the rainfall measurements, reservoir levels and gates opening. Forecasts are made in real-time, including inflows to the dam and discharges under different assumptions for gates operations during the inmediate future hours.