Abstract: Accurate short-term photovoltaic power interval probability prediction can effectively quantify the uncertainty of photovoltaic power prediction, which is very important for the operation and dispatch of new power systems to avoid risks. In order to improve the prediction performance of the model, a fuzzy C-means (FCM) clustering method is proposed based on the data characteristics of meteorological variables, and the historical data set is clustered into sunny days, sunny to cloudy and rainy days. The historical data of weather types are used as training samples to train the model. The convolutional neural network (CNN) model has excellent feature extraction advantages, and the bidirectional long short term memory (BiLSTM) neural network model is good at capturing long-term dependencies in long time series in both directions. A QR-CNN-BiLSTM deep learning fusion model is proposed after making full use of the advantages of CNN model and BiLSTM model and the advantage of quantile regression (QR) model that can generate interval prediction results. Meanwhile, by taking into account of the various meteorological factors obtained by screening, photovoltaic power is predicted by fine time granularity classification interval at 5 manutes intervals. Finally, the probability density prediction results are obtained by kernel density estimation using cross-validation and grid search methods. Moreover, a variety of evaluation indicators are selected to evaluate the proposed model, and compared with the prediction results of the QR-LSTM and QR-BiLSTM models. The results show that: 1) FCM algorithm can effectively realize the clustering of photovoltaic historical data sets; 2) QR -The CNN-BiLSTM fusion model can generate high-quality interval prediction results at intervals of 5 minutes, and the average value of the comprehensive evaluation index of the 95% confidence prediction interval is reduced from 0.137 1 and 0.128 8 of QR-LSTM and QR-BiLSTM to 0.097 1; 3) kernel density estimation based on cross-validation and grid search methods can generate reliable PV power probability density prediction results.