Water Level Prediction for Small Reservoirs via Machine Learning and Error Correction

Many small reservoirs lack adequate monitoring, making flood risk management a major challenge. This study aims to develop a comprehensive machine learning framework for short-term water level prediction, comprising feature selection, deep learning algorithms, a model transfer strategy, and an error correction technique, and applies it to four representative small reservoirs in Jiangsu Province, including two treated as ungauged scenarios. The results show that historical water levels are the dominant predictors for short-term reservoir water level forecasting, whereas precipitation becomes increasingly important during flood events. Based on the selected predictor combinations, both LSTM and GRU achieved high accuracy in routine water level prediction at the four small reservoirs. Under the scheme without precipitation input, both models performed well, with LSTM showing a relative advantage in flood event simulation. After incorporating the precipitation factor, GRU showed greater overall robustness and generally outperformed LSTM, particularly under complex flood conditions. In addition, comparing multisource precipitation products indicates that GMCP-driven simulations were generally more accurate than those based on ERA5-Land. Specifically, the LSTM–GMCP scheme performed best for the Yuwa Reservoir, whereas the GRU–GMCP scheme performed best for the Wushanqian Reservoir. Overall, the proposed framework demonstrates strong potential for short-term water level forecasting and flood-process simulation in small reservoirs, and provides useful support for hydrological prediction in data-scarce regions.