To address the key challenges of insufficient comprehensive extraction and fusion of meteorological conditions, temporal features, and periodic characteristics of power in short-term power prediction of distributed photovoltaic (PV) stations, a TPE-CBiGRU model prediction method based on multi-scale feature fusion is proposed. Firstly, multi-scale feature fusion of meteorological features, temporal features, and hidden periodic features in PV power is performed to construct model input features. Secondly, CNN and Bi-GRU are utilized to model the feature relationships between PV power and its influencing factors from spatial and temporal scales, respectively, and the spatial-temporal features extracted are fused through an Add network. Finally, the Bayesian hyperparameter optimization method is adopted to further optimize network parameters, achieving the prediction of single-station PV power. Validation using measured data from a certain PV station shows that the proposed method enhances the comprehensiveness of feature information extraction from both data and model layers, significantly improving the accuracy of short-term PV power prediction. Compared with other prediction models, the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) are reduced by 26.11% and 35.64%, respectively, and the R-squared (R2) is increased by 3.07%.