Interpretable Benchmarking of PSO-Tuned Random Forest for Solar Irradiance Prediction Marginal Gains and Battery-Voltage Leakage Risk

This repository/dataset accompanies a study which compared a baseline Random Forest (RF) with a Particle Swarm Optimization-tuned RF (RF-PSO) using 4,637 ground-based observations of rainfall, air temperature, relative humidity, wind speed, wind direction, barometric pressure, and battery voltage (ABV). An 80/20 train-test split and SHAP attribution were used to evaluate predictive performance and model behavior. SI exhibited substantial variability (SD = 0.26), extreme rainfall skewness (45.29), and generally weak linear predictor-response relationships (|r| < 0.13), except for ABV (r = 0.48). Baseline RF achieved R² values of 0.961 for training and 0.742 for testing. RF-PSO produced comparable test performance (R² = 0.743) while reducing the train-test gap from 0.219 to 0.215, indicating modest regularization rather than a meaningful accuracy gain. SHAP ranked ABV, relative humidity, and wind speed as the most influential predictors. PSO therefore provided limited improvement over the baseline, while ABV’s dominance indicates a potential leakage risk requiring ablation testing, walk-forward validation, and cross-site evaluation Keywords: Solar irradiance estimation; Random Forest; Particle Swarm Optimization; Paired bootstrap; Sahelian climate; Proxy leakage