A dynamic meteorological correlation integrated hybrid method for photovoltaic output forecasting

With the large-scale integration of photovoltaic (PV) systems into the power grid, accurate PV output forecasting is crucial to ensure the safe and stable operation of the grid. Mountainous PV plants face significant challenges in accurate forecasting due to complex and variable meteorological factors and unclear dynamic correlations between these factors and PV output. While ensuring forecasting accuracy, it is also necessary to consider the computational costs encountered in practical engineering deployments. Based on this, we propose a dynamic meteorological correlation integrated hybrid method. First, through correlation analysis, dominant meteorological factors are identified to achieve computational dimensionality reduction, quantify the correlation strength between meteorological factors and PV output, and explore their dynamic correlation rules. Then, these dynamic correlation rules are integrated into the hybrid forecasting model's loss function, and an alternating training approach is adopted to realize collaborative training between the temporal identification module and XGBoost. Finally, the hybrid method is validated and evaluated at a PV plant in the Hengduan Mountains. Compared with baseline models, results show that under the complex and variable meteorological conditions of the abundant wet season, dry season, and normal season, proposed method achieves R² values above 0.95 for 7-day and 10-day forecasting horizons, with RMSE reductions ranging from 10 % to 30 %. This demonstrates the excellent forecasting accuracy of the hybrid model and provides a valuable reference for improving PV output forecasting accuracy in other regions with complex meteorological factors.