Data-driven forecasting of aeolian dust impacts on solar irradiance and PV output over the Eastern Mediterranean

Abstract Dust outbreaks cause abrupt, hard-to-anticipate drops in surface irradiance and photovoltaic (PV) generation, yet most operational irradiance products lack the counterfactual dust-free signal needed to quantify dust-driven energy losses. We develop a two-model Extreme Gradient Boosting (XGBoost) framework, driven by Copernicus Atmosphere Monitoring Service (CAMS) forecast fields, to emulate all-sky hourly irradiance and diagnose the radiative impact of mineral dust over the East Mediterranean. An all-sky model is trained on 1.6 million hourly samples and evaluated against 2.6 million daytime hours of measurements constructed from 472 distributed PV plants, providing an hourly all-sky forecast product not directly available from CAMS. Against the PV proxy, XGBoost consistently reduces the positive systematic bias of CAMS reanalysis across dust regimes, while random-error skill is regime dependent. Under low-dust conditions (aerosol optical depth (AOD) < 0.2), RMSE is reduced by 3.8%; under extreme dust (AOD > 0.6), RMSE is reduced by 10.1%. Under moderate dust co-occurring with clouds (0.2 ≤ AOD ≤ 0.6), bias is reduced but MAE can increase relative to CAMS. A complementary clear-sky model is used in counterfactual dust-free mode to derive a diagnostic dust-effect metric linking dust-induced irradiance reductions to PV losses, with a monotonic response confirmed against AERONET AOD.