This study investigates the seasonal variability of meteorological feature importance in forecasting Global Horizontal Irradiance (GHI) using machine learning (ML) and deep learning models. High-resolution solar and meteorological data from NREL’s NSRDB (24.25°N, 45.34°E, 740 m) were seasonally partitioned into winter, spring, summer, and autumn. Feature selection was conducted using Pearson correlation (threshold 0.25), followed by dimensionality reduction through Principal Component Analysis (PCA). Six ML models XGBoost, LightGBM, Random Forest, SVR, MLP, and LSTM were trained on the processed datasets, and SHAP analysis was used to interpret feature contributions. The results revealed that clear-sky irradiance parameters (GHI, DNI, DHI) consistently dominated GHI prediction (correlation >0.95; SHAP >10⁻¹), while features like temperature and relative humidity varied across seasons. Wind direction, though weakly correlated, showed increased influence in winter. PCA enhanced model stability in spring and winter but slightly reduced accuracy during periods of high irradiance variability. Overall, XGBoost and Random Forest models provided the most accurate and reliable forecasts across seasons.
Sarker et al. (Mon,) studied this question.