Metaheuristic-Optimized Support Vector Regression for Daily Solar Radiation Forecasting in Eastern Indonesia

This study examines whether metaheuristic hyperparameter optimization improves support vector regression (SVR) for one-day-ahead forecasting of daily global horizontal irradiance in Makassar, Kupang, Ambon, and Jayapura, Eastern Indonesia. The study employs daily NASA POWER time series data for 2015–2024, using 2015–2022 for training and expanding-window temporal cross-validation and 2023–2024 for testing. SVR hyperparameters were tuned by particle swarm optimization (PSO), grey wolf optimizer (GWO), and whale optimization algorithm (WOA), each run ten times under an evaluation budget matched to grid search, and compared with a default SVR and a persistence forecast using the Diebold–Mariano test. The results show that all SVR variants significantly outperformed persistence (skill scores 0.13–0.28; test RMSE 0.74–1.30 kWh/m²/day). PSO, GWO, and WOA converged to nearly identical hyperparameters and accuracy, with no significant difference in stability. They reduced validation error by 0.2–0.7% relative to grid search, yet on the test set they were significantly better only in Jayapura, equivalent in Kupang, and significantly worse in Makassar and Ambon. The findings of this study suggest that the choice of metaheuristic matters less than the representativeness of the validation period, which here spanned a La Niña-to-El Niño transition.