Photovoltaic Operating-Point Prediction Using Optimised Machine Learning and Polynomial Regression

Accurate photovoltaic (PV) operating-point prediction supports energy-yield assessment, supervisory control, and digital-twin applications, but optimization benefits can be overstated when smooth model-generated data are used. This study benchmarks six standalone metaheuristics and equal-budget random search for tuning support-vector regression (SVR), XGBoost, and multilayer perceptron models that predict maximum power, maximum-power-point voltage and current, open-circuit voltage, and short-circuit current. The public dataset contains 496 model-generated operating records. Model selection used temperature-grouped development validation, while six complete temperature groups were withheld for external-condition evaluation. Each model-search-seed run received 500 candidate evaluations. The selected DE-SVR model achieved an external mean standardized RMSE of 0.0102 and mean R2 of 0.9999. However, degree-two polynomial ridge regression was both simpler and more accurate, with an external mean standardized RMSE of 0.00485. Recomputed run-level Friedman tests found no Holm-significant pairwise optimizer contrasts within any regressor family. The models reproduced the source single-diode temperature response: the maximum-power coefficient was −0.461%/°C in the source records, −0.447%/°C for DE-SVR, and −0.461%/°C for polynomial ridge. Because all partial shading records occur only at 1000 W/m2, these results do not establish performance for partial shading under cloudy or lower-irradiance operating profiles. Generalization was weakest for an entirely unseen 50% shading severity, where current-related R2 values became negative. Population-based metaheuristics therefore offered limited practical value on this smooth response surface. Polynomial regression is the appropriate first-line reference, and measured multi-module field validation remains necessary before deployment.