Physics-Guided Aerosol-Aware Probabilistic Photovoltaic Forecasting: Multi-Country Hindcasts and Measured Transfer

Photovoltaic (PV) forecasting in aerosol-affected, data-scarce environments is challenged by atmospheric heterogeneity, limited local measurements, and forecast uncertainty. This study develops a physics-guided artificial intelligence framework combining issue time meteorological and aerosol predictors with deterministic solar geometry and clear-sky information while preventing target time atmospheric predictor leakage. A retrospective protocol was evaluated at seven sites in Iraq, Saudi Arabia, the United Arab Emirates, Egypt, India, Australia, and Namibia, using 2023 for training, 2024 for calibration, and 2025 for independent testing at 1, 3, 6, and 24 h horizons. The multi-country target is a normalized pvlib AC reference; a separate PVDAQ experiment provides measured validation under data scarcity. Across 28 site–horizon cases, the causal model reduced RMSE by 56.9% relative to smart persistence. Aerosol predictors yielded a mean incremental RMSE improvement of 0.49%, with significant gains in three of 28 comparisons and no significant degradation. Conformal calibration increased the mean P10–P90 coverage from 69.8% to 80.3%. Measured cross-site transfer reduced the mean RMSE by 15.2% with 7 days and 9.4% with 30 days of target data, while target-only learning was slightly better with a full year. Physics guidance provides a robust forecasting backbone, whereas aerosol information and transfer learning provide conditional value.