Multi-step reservoir inflow prediction using a rolling window strategy and decomposed LSTM

Abstract

Effective management of multi-purpose reservoirs requires precise planning and accurate data to balance competing objectives and constraints. Reservoir inflow forecasting is critical in this process, with deep learning models increasingly applied across various time scales, from hourly to annual predictions. This study integrated a two-layer stacked long short-term memory network with decomposed data and a rolling window technique to enhance multi-day reservoir inflow forecasting accuracy. The proposed framework was applied to the Lam Takhong Dam in northeastern Thailand, a tropical monsoon region characterized by distinct wet and dry seasons. The dataset included daily reservoir inflow, river discharge, and average rainfall records spanning multiple years. Four forecasting strategies were compared for up to 7-d predictions: multi-step prediction, rolling prediction, multi-step prediction with decomposition, and rolling prediction with decomposition. The results indicated that while all models performed similarly for short-term predictions, accuracy declined over longer forecasting horizons. The rolling window approach with decomposition consistently outperformed others, achieving an average correlation coefficient of 0.92 and an average Nash–Sutcliffe model efficiency coefficient of 0.78 at the 7-d forecasting horizon. These findings demonstrate the practical advantages of integrating decomposition into a dynamic forecasting framework, particularly in reducing error accumulation in extended hydrological predictions.