Aerosol Features Add Day-Ahead Predictive Information Beyond All-Sky Irradiance: A Modeled Desert Photovoltaic Forecasting Benchmark for Oman

Solar photovoltaic capacity is expanding rapidly across Oman and the wider Gulf, yet airborne dust remains the largest and least quantified source of generation variability in desert climates. Widely used machine learning forecasting pipelines are developed mostly on temperate sites and typically omit aerosol inputs. This paper asks whether explicit aerosol features, namely aerosol optical depth, PM2.5, and PM10, improve day-ahead forecasts at a desert site in Oman beyond what irradiance already captures. Because no open Omani PV dataset exists, one was assembled for Ibri (23.05° N, 56.42° E) from free global products: NASA POWER irradiance and weather, CAMS reanalysis aerosols, and a pvlib-modeled generation target over 2022 to 2024. A smart persistence baseline, CatBoost, and a stacked gated recurrent unit (GRU) were each trained with and without aerosol features. Aerosol features raised day-ahead skill by 6.08 percentage points for the GRU, a gain a paired bootstrap resolves from zero, and by 1.21 for CatBoost, where it does not; aerosol-aware CatBoost was the most accurate variant, but contrary to expectation the gain did not concentrate on high-dust hours. Beyond the site, the study is a worked example of a region building its own benchmark rather than importing one: the feature sets that dominate public PV pipelines encode a climatology that is not universal, and testing that assumption here cost four columns of free reanalysis data.