Deep learning models based on BP neural networks are widely used in classification and regression problems. However, they have a wide range of parameters and insufficient local and uncertain information representations. Granular computing has great uncertainty characterization capability and allows input patterns to be abstracted at higher levels. A parallel granular BP neural network is proposed by combining clustered granulation to reconfigure the BP neural network. First, the optimal samples are defined and used as the references for granulating. Features are granulated into granules. And samples are granulated into granular vectors. Secondly, the granule vector is input to the granule BP neural network, and the granules for classification are output after the granule BP forward calculation, and are reduced to the original classification results by the threshing process. The granular BP neural network defines granular activation functions and a granular loss function to fit the input pattern of granular vectors while retaining the original fully connected neural network structure. The problem to be processed is extended into granules by granulation, and the structured characteristics of the granules make the granular BP neural network computable in parallel. Finally, compared with DNN, SVM, DT and NB, PGNN has better classification accuracy and improved computational efficiency and enhanced generalization performance.