UA-Solar-Resource-Forecast: an open ERA5-based dataset for an ex-post solar PV-potential forecasting experiment over southern Ukraine (PV potential is a derived proxy, not measured generation)

An open dataset for an ex-post forecasting experiment on the solar PV generation potential over the main solar regions of southern Ukraine (Kherson, Mykolaiv, Zaporizhzhia, Odesa). Scope: (1) the target is a derived proxy — temperature-corrected ERA5 shortwave irradiance — and is NOT measured photovoltaic-plant generation (Ukraine publishes no actual RES generation); (2) it is an ex-post setup — both the target and the weather covariates (cloud cover, temperature) come from the Open-Meteo ERA5 reanalysis archive (CC-BY 4.0), whereas an operational day-ahead forecast would use a forecast of those covariates (not used here). The covariates are cloud cover and temperature, not the irradiance that defines the target, so the task is not trivially circular. Two versions: base_v1 (5 solar-geometry features) and improved_v2 (9 features). Every model improves; the best (MLP) reaches R² 0.977 / MAE 18 (vs 0.935 / 33 for the geometry baseline). Note: the high R² (~0.98) mainly reflects the strong daily/seasonal regularity and autocorrelation of solar potential (an easy-to-predict target), not a hard forecast — a solar-geometry-only baseline already reaches R² ~0.93; the practical gain from the covariates is the MAE reduction under cloudy regimes, not the high R² itself. 35,017 hourly rows (2021–2024). Fully open (no API token). Supplements the author's renewable-generation forecasting work (KMOSS-2025).