A two-stage hybrid framework for photovoltaic power forecasting using MTF-MTS-Mixer and physics-informed Bayesian neural networks

Accurate photovoltaic (PV) power forecasting is crucial for ensuring reliable power system operation under high renewable energy penetration. This study proposes a novel two-stage hybrid PV power forecasting framework that combines the MTF-MTS-Mixer for deterministic day-ahead forecasting and the Physics-Informed Bayesian Neural Networks (PI-BNN) for probabilistic intra-day forecasting. The framework integrates multi-scale temporal features (MTF), temporal–channel mixing, physics-informed modeling, and Bayesian uncertainty quantification. Using a two-year dataset from seven PV power stations, extensive experiments are conducted to evaluate forecasting accuracy, robustness, and uncertainty representation. For deterministic forecasting, the proposed MTF-MTS-Mixer leverages MTF and temporal–channel mixing operations, achieving superior performance across all stations. It attains an average correlation coefficient (CC) of 0.811 and reduces NRMSE and NMAE by 14.12% and 16.43%, respectively, compared with TCN. It also maintains relatively favorable performance under different weather conditions, particularly rainy scenarios. For probabilistic intraday forecasting, the PI-BNN incorporates real-time meteorological inputs and physical guidance, enhancing forecasting performance and uncertainty estimation. It increases the average CC by 0.141 compared with day-ahead forecasting. In terms of probabilistic evaluation, it achieves a 40.29% improvement in CRPS compared with DeepAR and produces better-calibrated prediction intervals, with reduced PINAW and PICP closer to the nominal confidence level. In addition, the PI-BNN shows improved reliability in uncertainty estimation across different weather conditions. Overall, the proposed framework demonstrates competitive and stable forecasting performance across multiple PV stations and weather conditions. It provides useful information for PV operation and scheduling under uncertainty, supporting more reliable decision-making in renewable-integrated power systems.