Dynamic confidence fusion of calibrated physical and multi-scale deep learning models for short-term photovoltaic power forecasting

Accurate short-term photovoltaic (PV) power forecasting is essential for renewable energy integration and grid dispatch. Physics-based models provide interpretable estimates based on irradiance conversion and temperature effects, but they often suffer from site-specific bias under complex meteorological conditions. Data-driven models can capture nonlinear temporal patterns, but they may generate physically inconsistent forecasts under low-irradiance, rapidly changing, or nighttime conditions. This study proposes PhysFusionNet, a reliability-aware physics-data fusion framework for 15-min-ahead PV power forecasting. The framework consists of a calibrated physical estimation branch, a physics-enhanced multi-scale convolutional neural network–bidirectional long short-term memory temporal learning branch, and an environment-aware dynamic fusion strategy. The physical branch provides an interpretable baseline using irradiance conversion, temperature correction, wind-cooling adjustment, and irradiance-regime-based calibration. The data-driven branch learns nonlinear temporal residual patterns from meteorological, temporal, physical, interaction, and historical statistical features. The fusion strategy adaptively balances the two branches using irradiance, time, temperature, weather-stability, and validation-error information. Experiments are conducted using 1 year of 15-min operational data from a monitored PV sub-array in a cold-climate grid-connected PV station. The dataset metadata, preprocessing protocol, chronological data split, model configuration, ablation experiments, parameter sensitivity analysis, and statistical significance testing are reported to improve reproducibility. The results show that PhysFusionNet achieves lower forecasting errors than the selected baselines on the studied dataset and reduces physically infeasible forecasts under low-power and nighttime conditions.