DOI: 10.62520/fujece.1994408 ISSN: 2822-2881

HelioDA-XFE: ERA5-Based Solar Radiation Estimation and PV Potential Classification over Eastern Anatolia, Türkiye

Ertuğrul Gül
Show PDF Cite
Accurate assessment of solar energy potential is a core input for renewable energy planning in high-altitude regions such as Eastern Anatolia, where ground measurement stations are sparse. This study proposes HelioDA-XFE, an explainable feature engineering framework that estimates daily surface solar radiation (SSRD) from same-day ERA5 reanalysis fields (concurrent estimation, not forecasting) and assigns relative, radiation-based photovoltaic (PV) resource classes over an analysis domain centred on the Eastern Anatolia Region of Türkiye. The framework consists of four steps: (1) feature generation based on lags, rolling means, and seasonality encoding, (2) feature selection with mutual information (MI), (3) XGBoost models for regression and for a three-class classification of relative PV resource classes defined by SSRD terciles, and (4) SHAP-based interpretability analysis. A total of 5,228,990 grid-cell days covering 2000-2025 on a rectangular 19x29 cell (551-cell) domain at 0.25° resolution were evaluated with an expanding-window cross-validation scheme that prevents temporal leakage. The model reached R² = 0.972 ± 0.002 and RMSE = 0.354 ± 0.009 kWh/m²/day in regression, and 92.8 ± 0.5% accuracy with 0.928 ± 0.005 macro-F1 in classification, outperforming linear and decision tree baselines in both tasks. SHAP analysis indicated that clear-sky radiation is the dominant driver (mean |SHAP| = 1.51 kWh/m²/day), followed by cloud cover layers as the second largest driver group, confirming the physical consistency of the model. A theoretical, temperature-corrected PV potential that excludes system losses yielded annual specific yields of 1799 kWh/kWp for the domain and 1824 kWh/kWp for Elazığ; transferring the trained model to the 0.1° ERA5-Land grid mapped the intra-provincial distribution of PV yield across 94 points within Elazığ province. Because the target is the ERA5 radiation field itself, these results describe temporal generalization within the ERA5 climatology rather than independent radiation accuracy, and the outputs are intended for relative comparison and pre-feasibility screening. Within that scope, the framework offers a computationally light, transparent, and locally trained tool for regional solar energy assessment.

More from our Archive