Towards an Agentic and Explainable Runtime Architecture for PV Forecast Delivery

Most Photovoltaic (PV) forecasting studies optimize prediction accuracy, but production systems also need safeguards and explanations that operators can inspect. Rather than proposing a new forecasting model, this paper introduces an agentic wrapper for day-ahead PV forecast delivery. The wrapper orchestrates a baseline predictor, applies runtime integrity checks, and corrects or flags outputs when basic physical and operational constraints are violated (e.g., nighttime consistency and capacity bounds). Each forecast is accompanied by structured explanation artifacts, such as triggered checks and reason codes, designed for auditability and integration with operational pipelines. Experiments on real PV plant data over a 24-hour horizon show that the proposed architecture matches the baseline accuracy while improving physical consistency and traceable behavior at runtime.