In order to capture the hidden structure of networks and reduce the dependence of community detection models on the choice of initial parameters, this paper proposes a community detection model based on hierarchical graph clustering networks with probability distribution (HGCPD). Firstly, graph convolutional networks are utilized to learn and cache feature representations of nodes in the graph. Then, a hierarchical clustering method based on node pair similarity probability is introduced to recursively construct community structures at different levels. Finally, the problem of model hyperparameter optimization is explored, and the Bayesian optimization methods are designed to automatically adjust parameters, thereby improving the efficiency of the model. Experiments on multiple network datasets of different scales show that the HGCPD model is superior to mainstream methods in terms of the accuracy and effectiveness of community detection, and the model's interpretability is verified through visualization.