Scenario-adaptive robust crossformer for short-term photovoltaic power forecasting under extreme weather

Extreme weather events increase the uncertainty of photovoltaic generation and create operational risks for power systems with high renewable-energy penetration. This study proposes a short-term PV power forecasting framework that combines Scenario-adaptive Robust Principal Component Analysis (SA-RPCA) with Crossformer for high-temperature, prolonged-rainy, and sandstorm conditions. The SA-RPCA module uses weather-regime-specific physical constraints to separate low-rank weather-power patterns from sparse abnormal disturbances, thereby improving the robustness and physical consistency of the forecasting inputs. The refined features are then fed into a Crossformer architecture that uses dimension-segment-wise embedding and two-stage attention to model temporal dependencies and cross-dimensional interactions among historical PV output and meteorological input variables, while forecasting future PV power. Experiments show that the proposed framework consistently outperforms Transformer, Informer, and Autoformer benchmarks. The largest gains are observed under extreme-weather regimes, where the model reduces forecasting errors while maintaining stable performance under rapid irradiance attenuation and weather-induced outliers. The measured sub-millisecond latency per rolling window permits forecast updates within each 15-min operating interval. These forecasts could support intraday dispatch, reserve and storage scheduling, ramp-risk warnings, and curtailment management during extreme weather.