Accurate power forecasting of distributed photovoltaic (PV) is crucial for the safe and stable operation of power systems. To enhance the ability of distributed PV forecasting models to accurately match and identify temporal and spatial information from historical data, a hybrid forecasting model for distributed PV based on similar time period matching and graph modeling theory is proposed. Initially, considering the temporal correlation of distributed PV power data, the method of similar time period matching is utilized to identify the most critical power periods for prediction, and an improved Transformer model is proposed to extract temporal features of PV power. Secondly, in response to the spatial correlation of distributed PV output, a graph structure model of distributed PV is constructed based on sub-area division results, and a graph attention mechanism-based multi-layer bidirectional long short-term memory (MBLSTM) neural network model is established to extract spatial features of PV power. Finally, a spatiotemporal feature fusion mechanism for distributed PV power is proposed, which enhances the model's understanding and utilization of spatiotemporal information, and a short-term power prediction ensemble model for distributed PV is established. Experimental results indicate that the proposed ensemble model can effectively extract both temporal and spatial information of distributed PV power, demonstrating higher prediction accuracy compared to other models.