A multi-stage season-aware framework integrating view-factor irradiance and thermal dynamics for short-term bifacial photovoltaic power prediction

Bifacial photovoltaic (bPV) power output is difficult to accurately calculate and forecast because of the combined impacts of thermal effects, seasonal weather fluctuation, and rear-side irradiation. This study presents a combined physical modeling and forecasting framework for seasonal bPV power prediction by integrating rear-side irradiance estimation using the view factor technique, thermal and power modeling, along with data-driven forecasting techniques. A robustness metric based on the sensitivity index (SI), coefficient of variation (CV), and skill score (SS) was employed as a stability-aware criteria for season-specific model selection. Seasonal divisions were incorporated to evaluate the effect of parameter variation on power calculation. The calculated 5-min bPV power, along with meteorological parameters, is utilized as input for seasonal power forecasting. The results indicate that forecasting accuracy depends significantly on the season, and no single model consistently outperforms the others across all seasons. The multi-layer perceptron (MLP) model performs best during winter (R2 = 0.974), while Lasso and MLP demonstrate superior performance in summer (R2 ≈ 0.994). During the monsoon season, the performance of the models becomes more comparable (R2 ≈ 0.99), whereas the random forest regressor (RFR) performs relatively better, achieving the minimum RMSE (32.26 W) among the evaluated models in the post-monsoon season.