A machine learning approach for hydroelectric power forecasting

Sustainable energy solutions are essential for a cleaner and green future, and hydroelectric power responds to this need. To optimize the performance of hydroelectric plants, accurate hydropower generation forecasting is essential for various purposes, like capacity planning, decision-making, and anomaly detection. This paper aims to develop a one-step ahead monthly forecasting model for hydroelectric plants integrated in Water Distribution Systems (WDSs) using machine learning algorithms not previously used in this context. In contrast to existing literature that mainly concentrates on run-of-river or storage-reservoir-based systems, the proposed approach applies to all WDSs-integrated plants, whose primary purpose is the delivery of water to consumers. Among more than 500 models tested, the Temporal Convolutional Network algorithmarchitecture, with specific hyperparameters, achieved the best performance (Symmetric Mean Absolute Percentage Error equal to 5.202).