Prediction of hydroelectric power generation with machine learning and innovative combined deep learning techniques

Abstract

Dams provide energy production by the accumulation and storage of water. Therefore, changes in weather conditions directly affect production capacity and energy efficiency. While the amount of precipitation determines the circulation capacity of water resources, temperature affects the evaporation rate of water and thus water levels. Flow is one of the critical parameters required to determine the amount of water needed for energy production and to ensure efficient energy production. Within the scope of this study, energy production forecasting models have been established for the Alt & imath;nkaya Dam Basin, which has significant potential for hydroelectric energy production in Turkey. In addition to long-short-term memory (LSTM) and feed-forward neural network (FFNN) methods, TPAFFNN-LSTM, which combines these methods with an innovative temporal pattern attention (TPA) mechanism, was also used. Random forest (RF) and extreme gradient boosting (XGB) are also used to evaluate the efficiency and accuracy of the proposed models. As a feature selection method, LASSO regression was applied before the analyses. Shapley Additive Explanations (SHAP) and Regression Receiver Operating Characteristic (RROC) analyses were used in the evaluation phase of all models. According to the results obtained, the nRMSE and NSE criteria of the TPAFFNN-LSTM method were obtained as 0.16 and 0.69, respectively. These results were found to be 18% and 19% more successful than the other methods. The proposed method represents a significant advancement in hydropower energy generation forecasting, providing a robust framework that combines depth of analysis with clarity of insights.