This paper proposes a short-term forecasting method for photovoltaic （PV） power generation tailored to different weather types， which leverages physical information to guide the model in achieving appropriate feature attention. First， historical data is categorized into distinct weather types using a clustering method. Based on the physical mechanisms of PV power generation， key features influencing power output under various weather conditions are analyzed， and corresponding forecasting sub-models are trained for each weather type. Next， an attention optimization mechanism is constructed， where physical information is used to adjust the attention weights of key features in each sub-model， thereby optimizing the learning patterns of the model under different meteorological conditions. Finally， experimental results on public datasets demonstrate that， compared to traditional forecasting models， the proposed method not only significantly improves forecasting accuracy but also enhances model interpretability and controllability， effectively increasing the reliability of PV power forecasting models in practical applications.