Accurate short term power forecasting across multiple wind farms under extreme weather remains challenging because of temporal misalignment between meteorological disturbances and power responses, heterogeneous coupling among meteorological factors, and the limited adaptability of fixed graph topologies. To address these challenges, this paper proposes a spatiotemporal forecasting framework that incorporates physical constraints, meteorological propagation, and multifactor coupling. First, physical operating constraints and temporal trend similarity are integrated to identify abnormal power periods and generate reliable anomaly labels. Second, an anomaly guided self attention module is developed to quantify dynamic interactions among meteorological variables and transform the high dimensional meteorological matrix into a unified meteorological impact feature. Third, time delay attention and a dynamic graph derived from the physical diffusion process are incorporated into a spatiotemporal graph neural network to capture delayed power responses and changing spatial dependencies. Experiments were conducted on 100 wind farms in Western Inner Mongolia. The proposed method achieves a normalized mean absolute error (NMAE) of 5.785% and a normalized root mean square error (NRMSE) of 7.542%, outperforming the comparison models. The results confirm its effectiveness and robustness under extreme weather conditions.