Hybrid Multi-Branch Attention–CNN–BiLSTM Forecast Model for Reservoir Capacities of Pumped Storage Hydropower Plant

Pumped storage hydropower plants are important resources for scheduling urban energy storage, which realize the conversion of electric energy through upper and lower reservoir capacities. Dynamic forecasting of reservoir capacities is crucial for scheduling pumped storage and maximizing the economic benefits of pumped storage hydropower plants. In this work, a hybrid forecast network is proposed for both the upper and lower reservoir capacities of a pumped storage hydropower plant. A bidirectional long- and short-term memory network (BiLSTM) is designed as the baseline for the prediction model. A convolutional neural network (CNN) and Squeeze-and-Excitation (SE) attention mechanism are designed to extract local features from raw time series data to capture short-term dependencies. In order to better distinguish the effects of different data types on the reservoir capacity, the correlation between data and reservoir capacity is analyzed using the Spearman coefficient, and a multi-branch forecast model is established based on the correlation. A fusion module is designed to weight and fuse the branch prediction results to obtain the final reservoir capacities forecast model, namely, Multi-Branch Attention–CNN–BiLSTM. The experimental results show that the proposed model exhibits better forecast accuracy in forecasting the reservoir capacity compared with existing methods. Compared with BiLSTM, the M A P E of the forecast values of the reservoir capacities of the upper and lower reservoirs decreased by 1.93 % and 2.2484 % , the R M S E decreased by 16.9887 m 3 and 14.2903 m 3 , and the R 2 increased by 0.1278 and 0.1276 , respectively.