Short-term hydropower generation prediction model for run-of-river hydropower plants considering hydrometeorological factors

Abstract

Run-of-river hydropower plants, with limited storage capacity, primarily relying on the variability of incoming flow. However, the intricate fluctuations in external hydrological and meteorological factors engender a robust non-linear interdependency between streamflow patterns and hydropower generation. This study treats short-term hydropower generation prediction as a multivariate time series task and proposes a novel hybrid deep learning model, named Dual Attention Mechanism-Convolutional Neural Network-Bidirectional Gated Recurrent Unit (DAC-BiGRU). The model was validated and evaluated using both hydrometeorological reanalysis data and hydropower generation data. The results demonstrate that superior performance of the DAC-BiGRU model compared to baseline models such as Long Short-Term Memory (LSTM), CNN-LSTM, CNN-GRU, and Support Vector Machine (SVM), with an 8.8 % reduction in Root Mean Squared Error (RMSE). The integration of streamflow and soil temperature as supplementary input variables enhances the generalization capacity and predictive accuracy of the DAC-BiGRU model. The simplicity and efficiency of the DAC-BiGRU model make it a novel and effective solution with significant engineering relevance in short-term hydropower generation prediction.