Accurate photovoltaic (PV) power forecasting is critical for grid integration, particularly in tropical regions with limited measurement infrastructure. This paper proposes a hybrid framework that combines a metaheuristically optimized physical model with machine learning (ML) regression to enhance PV power forecasting. Seven regression models, including linear, kernel-based, neural-network, and tree-ensemble methods, were evaluated using data from a 2.4 MW grid-connected PV plant in Cuba under two experimental setups: measured power (Case 1) and physically modeled power (Case 2). Performance was assessed using root-mean-square error (RMSE), mean absolute error (MAE), coefficient of determination (R2), mean absolute percentage error (MAPE), normalized RMSE (nRMSE), normalized MAE (nMAE), and mean bias error (MBE) under a hold-out scheme and fivefold cross-validation. Results show that artificial neural networks (ANN) and gradient boosting machines (GBM) achieved the highest accuracy, with Case 2 yielding R2 > 0.99 for 11 of the 14 model configurations. The main contribution is the use of a metaheuristically optimized thermal model to generate physically consistent target power values, which supports accurate forecasting in sensor-constrained tropical settings and shows good potential for adaptation to other PV technologies and climates.
