SCADA-Based AC Power Forecasting in a Utility-Scale PV Plant: Explainable HistGradientBoosting and Thermal Derating Quantification

Accurate photovoltaic (PV) power modeling is essential for grid operation, short-term scheduling and plant-level decision making. However, reported performance in the literature varies widely due to differences in feature selection, forecast horizon definition and evaluation protocols. In particular, true exogenous forecasting is often conflated with contemporaneous power tracking, leading to optimistic accuracy estimates with limited operational relevance. This study aims to clarify how feature availability, evaluation design and operational intent shape reported performance in PV power modeling. A leakage-aware and operationally realistic framework is proposed that explicitly distinguishes between exogenous forecasting and operational tracking tasks. Using high-frequency SCADA data from a utility-scale PV plant, two modeling scenarios are evaluated. Scenario A relies exclusively on exogenous meteorological inputs and represents a forecasting-oriented inference task, whereas Scenario B augments these inputs with synchronous electrical current measurements and is interpreted as operational tracking (nowcasting) rather than forward-looking prediction. Multiple regression models, with emphasis on tree-based ensemble methods, are evaluated under transparent and scenario-consistent train–test splitting strategies. Model behavior is examined using post-hoc explainability tools to assess physical consistency and temperature-induced power derating is quantified using a semiparametric regression approach. The results indicate that exogenous-only models recover physically meaningful relationships but exhibit higher numerical errors than current-augmented configurations, reflecting the informational advantage of synchronous measurements rather than improved forecasting capability. Rather than claiming superior numerical accuracy, this work demonstrates how modeling assumptions and evaluation choices directly influence reported performance and its interpretation. By aligning feature availability with operational intent, the proposed framework provides practical guidance for the development, evaluation and deployment of PV power models in real-world applications.