Multi-step-ahead reservoir inflow forecasting by artificial intelligence techniques

Abstract

Taiwan is located in themonsoon zone of the North Pacific Ocean and experiences an average of 4-5 typhoons annually. The particular topography of Taiwan makes rivers short and steep, and thus rivers flow rapidly from catchments to reservoirs within a few hours during typhoon events. This study aims to construct realtime multi-step-ahead reservoir inflow forecast models by using Artificial Neural Networks (ANNs) based on radar rainfall data and reservoir inflow data. The Back PropagationNeural Network (BPNN) and the Recurrent Neural Network (RNN) are adopted for forecasting. Results indicate that the correlation coefficients in the testing phases of both models exceed 0.86 for one- to three-hour-ahead forecasts and exceed 0.69 for six-hour-head forecasts. The RNN model outperforms the BPNN model, which indicates the recurrent property of the RNN can effectively improve forecast accuracy when making several step-ahead forecasts. Results demonstrates that the constructed multi-step-ahead rainfall-runoff models can provide valuable instantaneous inflow forecasts for the coming six hours so that decision makers can implement more suitable reservoir operations in consideration of inflow forecasts, rather than just depend on historical scenarios.