In order to stabilize the process which integrates photovoltaic (PV) power into a power grid, a hybrid deep learning model based on potential season category in the data is proposed after researching the prediction of PV output power. The overall model is divided into three stages, namely, clustering, training and prediction. In the clustering stage, correlation analysis and self-organizing mapping are employed to select the features with the highest correlation in historical data. In the training stage, the CNN, LSTM and the attention mechanism are combined to construct a hybrid deep learning forecasting model. In the prediction stage, a particular classification model is selected based on the month of the testing dataset. The experimental result shows that this proposed model has significantly improved prediction accuracy in terms of a time interval of 7.5 min. This work is supported by the Basic Public Welfare Research Project of Zhejiang Province (No. LGF18F020017).