PARA-PV: Physics-Aware AI Boosts Solar Power Forecasting Accuracy

Accurate forecasting of photovoltaic (PV) power is crucial for managing electricity grids and integrating renewable energy, but it's challenging due to weather variability, day-night cycles, and physical constraints. A new framework, PARA-PV (Physics-Aware Retrieval-Augmented PV Prediction), has been developed to address these complexities by embedding physical knowledge throughout the forecasting process. PARA-PV first encodes PV observations into representations and then uses a physics-aware retrieval-augmented learner to find historical data and trajectories that match the current conditions. This provides a physically grounded initial forecast. To enhance this, the base forecast is then calibrated against a frozen Chronos time-series foundation model using a lightweight adapter, allowing the system to incorporate general temporal regularities without overriding the physics-based prediction. Finally, a physics-aware distribution shift correction module adjusts the forecast for changing weather and diurnal regimes, applying targeted mean-shift and scale corrections. The framework also employs a physics-constrained loss function that adaptively reweights errors across different operational regimes (peak, ramping, night-time) to ensure critical states are learned effectively.