Multi-model validation and site-specific seasonal bias correction of rooftop photovoltaic energy yield

Accurate prediction of photovoltaic (PV) energy yield remains challenging under strong seasonal variability. This study evaluates PVGIS, the System Advisor Model (SAM), and an hourly physics-based analytical model against monthly inverter-recorded AC energy from a 111.93 kWp rooftop installation in Bacău, Romania. The dataset covers 2023–2024 and the period from 1 January to 9 December 2025. The analytical model and SAM used year-specific Solcast meteorological inputs, whereas PVGIS retained its SARAH-3 climatology and native calculation chain. The analytical–Solcast chain achieved the lowest overall RMSE (685 kWh/month), followed by SAM–Solcast (993 kWh/month) and PVGIS (1,309 kWh/month). Its winter nRMSE was also lowest at 18.3%, compared with 28.3% for SAM and 28.7% for PVGIS. These results characterize complete modelling chains and do not isolate the respective effects of meteorological inputs and model architecture. Winter deviations could not be attributed mechanistically without on-site irradiance and snow-coverage measurements. Leave-one-year-out analysis showed that a global bias coefficient reduced mean validation RMSE by 3.5%, whereas seasonal correction reduced it by 17.4%, from 683.30 to 564.45 kWh/month. The uncorrected analytical–SAM ensemble provided no consistent improvement. Ensemble weighting after model-specific seasonal correction yielded the lowest mean RMSE (473.40 kWh/month), 30.7% below the uncorrected analytical baseline, while ensembles including PVGIS provided no further improvement. Because the analysis comprised only three annual folds from one installation, the calibrated factors, weights, and observed improvements remain site-specific and require longer-term, multi-site validation.