Optimized Machine Learning Framework for Predicting Photovoltaic Soiling in Arid Desert Climates

Soiling is a major source of photovoltaic (PV) energy loss in desert environments, where day‐to‐day dust accumulation alternates between steady deposition, rapid storm‐driven degradation, and rare rain‐driven recovery. This study evaluates a light gradient boosting machine (LightGBM [LGBM]) quantile‐regression framework (Q‐LGBM) that issues 72‐h soiling‐ratio forecasts, comprising a conditional median, an 80% prediction interval, and a directional signal, for an outdoor test facility in Doha, Qatar. Inputs combine weather forecasts, air‐quality metrics, solar irradiance, and autoregressive soiling‐ratio lags. Expanding‐window cross‐validation and a sequential temporal holdout are benchmarked against linear‐decay and persistence baselines. Without rain postprocessing, Q‐LGBM reaches a holdout RMSE of 0.0092, statistically indistinguishable from linear decay (Diebold–Mariano p = 0.951), but detects soiling recovery markedly better (upward recall 0.773 vs. 0.227). Under cooler, dustier holdout inputs, the 80% interval widened from 0.040 to 0.060, and empirical coverage rose from 83.3% to 96.2%. Distribution shift diagnostics across inputs, soiling ratios (SRs), and calibration metrics show that input shift coincided with this interval widening, suggesting a candidate trigger for hyperparameter retuning. A supplementary rain‐reset rule yields the lowest cross‐validation RMSE (0.0083) but degrades on the holdout (0.0160) because of numerical precipitation forecast false alarms and subcleaning rainfall.