Exploring the role of the long short-term memory model in improving multi-step ahead reservoir inflow forecasting

Daily inflow forecasting is of vital importance in reservoir economic operation. In the context of hydrometeorological forecasting, the effectiveness of the data-driven models has been demonstrated as bias correctors for physically-based models or direct forecasting models. However, existing studies only highlight the performance improvements provided by the data-driven model, lacking a comprehensive investigation on whether the data-driven model should be used as bias correctors or direct forecasting models. This study constructs long short-term memory (LSTM)-based preprocessing and postprocessing techniques for a hydrological model, which are tested by linear scaling preprocessing and autoregressive (AR) postprocessing models. The integrated model is compared with the LSTM-only model. The Shuibuya and Zuojiang reservoirs in China are selected as case studies. Results indicate that: (1) LSTM-based bias correctors are effective in both preprocessing and postprocessing and (2) the integrated model is comparable to the LSTM-only model when trained with four or more years of data, while it is better than the LSTM-only model when trained with less data. These findings demonstrate that data-driven methods can effectively correct the bias in physically-based model output, and integrating the physical and data-driven models is useful in improving multi-step ahead reservoir inflow forecasting if limited data can be obtained.