An Explainable Agentic Architecture for Multi-Site Photovoltaic Power Forecasting

Abstract

Photovoltaic (PV) power forecasting is critical for grid stability and renewable energy integration. However, existing approaches prioritize predictive accuracy while neglecting operational requirements: physical validity under noisy data, transparent decision-making, and scalable multi-site deployment. This paper proposes an explainable architecture for multi-site photovoltaic forecasting that extends a validated hybrid ensemble pipeline through system-level design. Forecasting is structured as an explicit reasoning process comprising perception, inference, validation, explanation, and action stages. Each photovoltaic site uses an autonomous Forecasting Unit (FU), deployed as a microservice for scalable and consistent operation across sites. Explainability is achieved through explicit model orchestration, constraint-based validation, and traceable fallback mechanisms. Validation on real-world PV systems demonstrates that the architecture maintains competitive forecasting accuracy while ensuring physical validity and transparent model behavior. All forecasts are operationally applicable with traceable decision-making processes. These results highlight the relevance of architectural organization and an agentic-by-design (deterministic reasoning loop) architecture, based on explicit validation and orchestration mechanisms, for deploying trustworthy photovoltaic forecasting systems in industrial environments.