摘要
This study presents a comprehensive evaluation of nine forecasting models namely statistical, machine learning, deep learning, hybrid, and probabilistic for monthly average solar irradiance and wind speed prediction in South Africa. Models were trained using three historical windows (2003–2023, 2013–2023, 2018–2023) and assessed with MAE, MAPE, RMSE, and visual diagnostics. Paired t-tests, Wilcoxon tests, and Friedman post-hoc analyses were conducted to statistically validate differences, including comparisons against baseline Climatology and Seasonal Naive forecasts. Results show that LSTM and GBQR consistently outperformed other models across both domains, achieving the lowest errors and statistically significant improvements over baseline methods in most periods. For solar irradiance, LSTM achieved MAE of 5.50 W/m 2 and RMSE of 7.26 W/m 2 , while wind speed achieved a MAE of 0.34 m/s and RMSE of 0.42 m/s. Solar forecasting benefited from ∼ 10–11-year datasets, whereas wind forecasting was more sensitive to model architecture than training duration. Baseline methods were only competitive in periods of low variability, highlighting the advantage of models that capture temporal dependencies, stochastic variability, and seasonal patterns. Monthly forecasting also provides a computationally efficient framework for assessing model robustness and overfitting risk before high-resolution operational deployment. These findings confirm that models integrating temporal memory and probabilistic learning, such as LSTM and GBQR, offer the most reliable foundation for renewable energy planning, system sizing, and real-time operational forecasting. These results support future integration into hourly dispatch models to improve battery energy storage system operation and grid coordination in virtual power plant applications.