Abstract: SWith high-penetration renewable energy integration, the accuracy of photovoltaic power forecasting affects grid dispatch security. To address the spatial mismatch between NWP and plant micro-meteorology, and the inadequacy of traditional methods in characterizing continuous weather variations, a regional PV power forecasting method based on Spatio-Temporal State Representation and dual-path deviation correction is proposed. Lagged feedback, sliding-window statistics, and irradiance fluctuation features are introduced to enhance the description of meteorological dynamics. Combining physically constrained and data-driven dual-path deviation correction, an adaptive fusion mechanism improves forecast stability under complex weather. Experiments show that the proposed method outperforms comparison models under sunny, overcast, and cloudy conditions, with the root mean square error reduced from 468.0117 to 212.6154, demonstrating favorable accuracy and robustness.