Generalized versus Panel-specific Machine Learning Models for Photovoltaic Power Estimation Under Limited Multi-Module Field Data

Limited multi-module field data creates a trade-off between generalized models that pool observations from several photovoltaic modules and panel-specific models that preserve module dependent behavior but use fewer training samples. This study compares three modeling scopes for photovoltaic power estimation using 63 observations from three nominally 100 W modules measured over four days. Generalized-Shared, Generalized-Panel-Aware, and Panel-Specific scopes were evaluated using Ridge Regression, K-Nearest Neighbor Regression, Radial Basis Function Network Regression, and Random Forest Regression under nested leave-one-day-out validation. A datasheet physical power estimate was also evaluated as a non-machine-learning baseline. Panel-Specific Ridge achieved the best overall performance with an RMSE of 7.3236 W, MAE of 5.6750 W, and R2 of 0.8788, reducing RMSE by 60.23% relative to the physical baseline (18.4161 W). Generalized-Panel-Aware Random Forest achieved an RMSE of 7.7876 W overall and became the best configuration for P3 (6.3310 W) and for the Tuesday left-out day (4.9565 W). Generalized KNN and panel-aware RBF also benefited from pooled training data, whereas Panel-Specific Ridge remained the strongest regularized linear configuration. Wilcoxon signed-rank tests with Holm correction found no statistically significant scope differences across the 12 panel-day units; the largest scope effect was observed for panel-aware Random Forest (rank-biserial = -0.4872, adjusted p = 0.3027). The results show that the benefit of pooling or specialization is algorithm and operating condition dependent and should be interpreted within the limitations of the small single-site field dataset.