Real-time reservoir operation using recurrent neural networks and inflow forecast from a distributed hydrological model

Abstract

Large-scale reservoirs play an essential role in water resources management for agriculture irrigation, water supply and flood controls. However, we need robust reservoir operation systems under both normal flow and extreme flow conditions. In this study, we applied recurrent neural networks (RNN) to simulate the operation of three multi-purpose reservoirs located in the upper Chao Phraya River basin. Two reservoirs have the function of multiannual flow regulation and one has the function of incomplete annual regulation. The goal of this study is to explore the applicability of RNN models for operation of reservoirs with multiannual flow regulation under different flow regimes, especially under extreme floods and droughts. We used three RNNs, namely nonlinear autoregressive models with exogenous input (NARX), long short-term memory (LSTM) and genetic algorithm based NAXR (GA-NAXR) for reservoir operation based on historical data. For real-time water resources management, an accurate inflow forecast is required to provide a real-time reservoir outflow, and thus we also carried out a real-time reservoir operation using the RNN and the inflow forecast by a distributed hydrological model. Results show that (1) GA-NARX has the highest accuracy among three RNNs and is more stable than the original NARX by optimizing the initial conditions, although it takes longer training time than NARX and LSTM; (2) GA-NARX-based operation model is effective under extreme floods and droughts; and (3) the real-time operation system combining the GA-NARX and the distributed hydrological model has reasonable accuracy in both wet season and dry season. RNN-based operation model developed in this study has potential applicability in practical water management, and the model combining the hydrological prediction is specially useful for real-time reservoir operation.