Prediction of reservoir water levels via an improved attention mechanism based on CNN − LSTM

Abstract

Water level prediction is crucial for flood control scheduling and water resource management. The application of various deep learning methods to water level prediction in reservoirs is limited. Accurate water level prediction aids in optimizing reservoir operation strategies, ensuring flood safety downstream and meeting water supply demands. To achieve accurate predictions, a new structure based on a convolutional neural network − long short-term memory (CNN − LSTM) model is proposed, which incorporates a self-attention mechanism and a local attention mechanism in an SL − CNN − LSTM coupled model. Using the Three Gorges Reservoir head area in China as a case study, hydrometeorological data from three points in the reservoir's head area and upstream water level characteristics are used as input variables. Data collected every six hours from 2008 to 2021 were used, with the model trained and tested at an 8:2 ratio. The study revealed that a two-layer CNN configuration performed best in most models. The SL − CNN − LSTM-2 model achieved the best performance across all the metrics, with an R² of 0.9988, an MAE of 0.2767, an RMSE of 0.3404, and a MAPE of 0.1717, particularly for extreme water level predictions with minimal residuals, validating its strong ability to balance long- and short-term dependencies. Additionally, the model effectively extracts features and captures critical information in time series data, balancing learning capacity and computational efficiency. The research results are highly important for water resource management in large reservoirs, providing reliable technical support for flood control scheduling and water resource optimization.