Scenario-robust irradiance correction with multivariate meteorological features for short-term photovoltaic power forecasting under complex weather conditions

Photovoltaic power forecasting (PVPF) under complex weather remains challenging because numerical weather prediction (NWP) irradiance often contains large errors, while rapid irradiance fluctuations weaken the generalization capability of conventional models. Existing post-processing methods improve irradiance forecasts, but they still struggle to characterize mixed irradiance patterns and scenario heterogeneity. To address these limitations, this study proposes a scenario-aware deep learning framework for short-term PVPF. First, two representative complex-weather scenarios, sustained low-irradiance processes and strong irradiance ramp processes, are analyzed, and a cloud-wind composite disturbance index (CWCDI) is developed to characterize meteorological disturbances. Then, an irradiance correction model (GWNet-MoE) integrating Graph WaveNet (GWNet) and a mixture-of-experts mechanism (MoE) is proposed to achieve scenario-adaptive NWP irradiance correction under complex weather conditions. Graph WaveNet captures spatiotemporal dependencies among multiple PV stations, while the MoE layer learns scenario-specific correction patterns. Finally, the corrected irradiance and multi-source meteorological variables are fed into a channel-independent patch time series Transformer for PV power forecasting. Experimental results show that the proposed framework consistently outperforms comparison models under complex weather, achieving the average absolute root mean square error (RMSE) decrease of 0.0356 compared with Patch Time series Transformer (PatchTST) using original NWP irradiance.