Predicting Hydropower Production Using Deep Learning CNN-ANN Hybridized with Gaussian Process Regression and Salp Algorithm

The hydropower industry is one of the most important sources of clean energy. Predicting hydropower production is essential for the hydropower industry. This study introduces a hybrid deep learning model to predict hydropower production. Statistical methods are unsuitable for modeling hydropower production because their accuracy depends on seasonal and periodic fluctuations. For accurate predictions, deep learning models can capture daily, weekly, and monthly patterns. Since ANNs may not capture latent and nonlinear patterns, we use deep learning models to predict hydropower production. We used Convolutional Neural Network-Multilayer Perceptron-Gaussian Process Regression (CNNE-MUPE-GPRE) to extract key features and predict outcomes. The main advantages of the hybrid model are the quantification of production uncertainty, the accurate prediction of hydropower production, and the extraction of features from input data. We use a binary SSOA to select optimal input scenarios. The new model is benchmarked against the long short term memory neural network (LSTM), Bi directional LSTM (BI-LSTM), MUPE, GPRE, MUPE-GPRE, CNNE-GPRE, and CNNE-MUPE models. The models are used to predict 1-, 2-, and 3-day ahead power. The root mean square error values of CNNE-MUPE-GPRE, CNNE-MUPE, CNNE-GPRE, MUPE-GPRE, BI-LSTM, LSTM, CNNE, MUPE, GPRE were 578, 615, 832, 861, 914, 934, 1436, 1712, and 1954 KW at the 1-day prediction horizon. The RMSE of the CNNE-MUPE-GPRE was 595, 600, and 612 at the 1-day, 2-days, and 3-days prediction horizons. Extending the prediction horizon degrades accuracy. The uncertainty of the CNNE-MUPE-GPRE model was lower than that of the other models. The CNNE-MUPE-GPRE model is recommended for more accurate hydropower production predictions.