To enhance the accuracy and stability of photovoltaic (PV) power output forecasting under complex and highly variable meteorological conditions, a physics-data fusion-driven strategy is adopted, and a physical feature expansion ASReLU-CNN-LSTM method for short-term PV power forecasting is proposed. First, an improved solar trajectory model is used to dynamically correct the tilted surface irradiance so that it accurately reflects the actual irradiance received by PV modules. Subsequently, a PV conversion model and a lightweight feedforward network are employed to expand the dataset with relative power features. An adaptively smooth rectifier linear unit (ASReLU) is then designed, in which parameterized adaptive smoothing is introduced to enhance the negative-feature extraction capability of the convolutional neural network (CNN). Finally, the dataset augmented with physical features is fed into the ASReLU-CNN-LSTM model for PV power prediction. Experimental results on datasets from two distinct climatic regions demonstrate that the proposed method achieves high prediction accuracy and strong generalization capability.