Artificial intelligence-based optimization of a giga-scale bifacial photovoltaic power plant using partitioned random vector reinforcement learning and Eurasian-Lynx optimization

Photovoltaic (PV) systems are inherently intermittent due to weather-driven variability in solar irradiance and fluctuating load demands, which pose challenges for short-term forecasting and real-time energy management in power grids. This challenge is addressed by proposing an accurate PV power data-based prediction model using a robust AI framework combining Random Vector Functional Link (RVFL) with the Eurasian-Lynx Optimization Algorithm (ELOA). The developed RVFL-ELOA modeling was applied to predict the PV active power using historical datasets collected from the 1.50 GW Sudair solar PV park, Riyadh, Saudi Arabia, with a 15-minute resolution for an entire year. ELOA was utilized to identify the optimal selection of RVFL hyperparameters to maximize prediction accuracy. Moreover, a correlation approach was employed to select the critical input features that reduced the data dimensionality and improved the robustness of the models. In addition, the RVFL-ELOA was evaluated against standalone RVFL and multivariate adaptive regression splines (MARS) models via eight statistical indicators, with validation through monitoring and measurements of the PV power park. The statistical findings manifested the superiority of the proposed RVFL-ELOA method compared to the standalone RVFL and MARS for the performance prediction of the PV power plant. The results highlighted the critical role of integrating machine learning algorithms with hyperparameter optimization techniques, offering valuable insights into achieving adaptability and improving the smart-grid integration of PV power plants under diverse seasonal conditions.