Hybrid WRF–Machine Learning Irradiance Correction, POA Transposition, and PV Module Thermal Modeling for Photovoltaic Forecasting-Input Assessment and Monitoring

Abstract

Reliable photovoltaic (PV) monitoring requires meteorological inputs that remain interpretable through horizontal irradiance, module-plane irradiance, temperature, and electrical-output layers. This study evaluates a component-wise WRF–machine-learning–POA–thermal workflow using complementary field datasets: a synchronized 2021 WRF-ML/electrical dataset and an independent 2017 POA/GTI validation dataset. WRF-derived variables were treated as retrospective meteorological inputs for post-processing and forecasting-input assessment. Random Forest, Gradient Boosting, neural networks, mean bias-corrected WRF, and Ridge MOS baselines were tested for GHI correction; empirical, Perez, Hay–Davies, and isotropic models were compared for POA/GTI transposition; and five module temperature models were assessed. In a random 80/20 held-out test, Random Forest and Gradient Boosting reduced irradiance RMSE from 139.66 W/m2 for raw WRF to 75.92 and 75.96 W/m2, respectively. In blocked temporal validation, however, raw WRF was more stable for month-wise irradiance, and the physics-inspired Ridge baseline was more robust for leave-one-month-out AC/DC power prediction. Perez gave the best POA/GTI agreement, while NOCT and King/Sandia gave the lowest thermal errors. The results support a protocol-dependent, traceable input-chain assessment for PV monitoring and identify the calibration, metadata, and timestamp controls needed before operational power-forecasting claims can be generalized.