A Hybrid Deep Learning Model for Short-Term Hydropower Generation Prediction Incorporating Hydrometeorological Factors

Abstract

Abstract Hydropower is a clean and renewable source of energy, reliable forecasts of hydropower generation are critical for reservoir management and efficient use of water resources. In this study, a Double Attention mechanism-Convolutional Neural Network-Bidirectional Gated Recurrent Unit(DAC-BiGRU) hybrid deep learning model considering hydrometeorological factors is proposed for predicting short-term hydropower generation. DAC-BIGRU first extracts the local features of the input data using a 1-dimensional CNN and weights the features using an attentional mechanism instead of a pooling layer. Then passes features to the BiGRU layer for time series modelling. Finally performs the attentional mechanism weighting again and maps the results to the final output via a dense layer. The Qiamusa and Aertashi hydropower stations on the mainstem of the Yarkant River in Northwest China are used as the study object for cross-training and validation. The Pearson Correlation Coefficient and Maximum mutual Information Coefficient were used to determine hydropower generation, steamflow and soil temperature for the past 7 days as input features.Through rigorous and reasonable evaluation, the results demonstrate that: Root Mean Squared Error (RMSE) of DAC-BiGRU on the test set decreased by 8.8% on average compared with CNN-LSTM and CNN-GRU. DAC-BiGRU model performance was significantly better than CNN-LSTM, CNN-GRU, Long Short-Term Memory (LSTM) and Support Vector Machine (SVM). In addation, relative error of DAC-BiGRU mainly occurred from October to March, RMSE of the DAC-BiGRU decreased by about 3% after adding streamflow and soil temperature as factors. The prediction performance of DAC-BiGRU for the next 3–7 days began to decline significantly.