Cross-Plant Transfer Assessment of Photovoltaic Performance Using Data-Driven Models and a Normalized Technical Factor

Assessing whether data-driven photovoltaic (PV) models can generalize beyond the plants used for training remains a critical challenge when operational data are available for only a limited number of installations. This study proposes a methodological framework for cross-plant transfer assessment, focusing on spatial generalization rather than within-plant temporal fitting. Daily operational data from three utility-scale PV plants are combined with satellite-derived irradiance and meteorological variables to construct a normalized, dimensionless technical performance indicator, closely related to established PV normalization metrics and designed to reduce scale and site effects. Forecasting models—including decision-tree ensembles (Random Forest and gradient-boosted trees), multilayer perceptrons (MLP), and convolutional–recurrent neural networks (CNN–LSTM and CNN–GRU)—are evaluated by training on data pooled from two plants and testing exclusively on the third. Hyperparameters are optimized using Optuna under a unified tuning strategy. Performance is assessed using MAE, sMAPE, and RMSE, with emphasis on metrics appropriate for normalized targets. Results show that no single model dominates across all transfer scenarios and that forecasting performance degrades under spatial transfer in a plant-dependent manner, even after normalization. Diebold–Mariano tests against a Kriging-based spatial baseline provide strong statistical evidence that the learning-based models deliver predictive power beyond spatial smoothness alone across all evaluated plants. The contribution of this work lies in quantifying the limits of cross-plant generalization and in providing a statistically grounded framework for comparative PV performance assessment in data-scarce settings.