Transfer Learning-Enhanced Transformer for Robust Solar Irradiance Forecasting in Tropical Renewable Energy Integration

Accurate medium-range solar irradiance forecasting in tropical regions remains highly challenging due to rapid convective development, short-lived temporal dependencies, and substantial domain mismatch between observational and forecast-driven inputs. This study proposes a two-day-ahead forecasting framework that combines an encoder-only Transformer with a transfer-learning strategy linking ground-based observations to WRF-derived predictors. The model is pertained on high-resolution observational sequences to learn physically consistent temporal representations, and subsequently fine-tuned on WRF inputs to mitigate domain shift during operational deployment. Using full-year 2019 data from Thailand, the proposed Transformer+TL architecture achieves an RMSE of 119.15 W/m ² , RRMSE of 28.94 $$%$$ , and R²=0.8683 , outperforming a tuned XGBoost baseline with statistically significant gains. Ablation experiments confirm that self-attention and transfer learning contribute complementary improvements, while robustness analyses show stable performance across dry and monsoon regimes. Sensitivity evaluation further reveals a short predictive memory of approximately 1–2 h in tropical irradiance, with error increasing exponentially for longer lookback windows-consistent with the fast decorrelation of convective cloud systems. The resulting framework is accurate, scalable, and operationally relevant for photovoltaic scheduling, reserve management, and tropical grid reliability, providing a strong foundation for future probabilistic, multi-site, and physics-informed solar forecasting systems.