Intelligent forecasting of photovoltaic power output using integrated meteorological and operational data

Background Integrating solar energy into smart grids necessitates accurate forecasting models to ensure grid stability and efficient energy management. Traditional methods struggle with weather-induced fluctuations and system-specific variations, while existing data-driven approaches fail to capture the nonlinear relationships between meteorological conditions and photovoltaic (PV) operational parameters. Results We propose the Spectral-Aware Photovoltaic Encoder (SAPE), a physics-informed deep learning framework that integrates wavelength-resolved irradiance modeling with real-time operational data for PV power output forecasting. SAPE incorporates three key innovations: (i) spectral-resolved photon modeling that accounts for wavelength-dependent quantum efficiency variations, (ii) physics-constrained I-V learning that enforces single-diode model consistency during training, and (iii) a context-guided feature modulation mechanism that conditions predictions on installation-specific physical priors. Validation across four benchmark datasets (SKIPP’D, Solcast, NSRDB, and GEFCom2014) demonstrates consistent improvements, with reductions of up to 10.3% in RMSE and 10.9% in MAPE compared to the strongest baseline (iTransformer). The model achieves R² values ranging from 0.82 to 0.91 across the four datasets and maintains reasonably calibrated 90% prediction intervals (PICP ≥ 0.87). Conclusion SAPE demonstrates robust forecasting performance across diverse weather conditions, geographic regions, and forecast horizons (1–12 h ahead), offering a practical and generalizable solution for intelligent PV power management in smart grid systems.