Interpretable cloud cover prediction for dynamic solar prediction in low-data environments

Cloud cover strongly influences solar irradiance variability, directly affecting photovoltaic (PV) energy generation. Rapid fluctuations in cloud properties such as type, thickness, and movement can cause unpredictable drops in solar output, challenging grid reliability and energy dispatch. Accurate cloud cover prediction is often hindered by sparse meteorological data, particularly in geographically remote or sensor-deficient regions. To address this, we propose a framework that employs Gaussian Mixture Models (GMMs) to generate physically consistent synthetic meteorological data, augmenting limited training datasets. This approach is applied to tree-based machine learning models, including Random Forest, CatBoost, and XGBoost, with SHAP (SHapley Additive exPlanations) integrated to enhance interpretability. Experimental results show improved accuracy and robustness, with Random Forest achieving a Mean Absolute Error (MAE) of 11.7767 ± 0.1091, Root Mean Square Error (RMSE) of 17.2762 ± 0.2604, and R² of 0.8092 ± 0.0043. SHAP analyses reveal more stable feature contributions, particularly for dew point and relative humidity. This framework has significant practical value for solar forecasting in Southeast Asian regions with limited sensor networks, enabling accurate cloud cover prediction, improved grid reliability, and scalable edge deployment for solar energy integration.