Abstract: In order to improve the accuracy and convergence speed of attack source location in power communication network, an attack source location method based on convolutional neural network is proposed.The first-order spatial regression model is used to analyze the spatial structure characteristics of the target equipment of the power communication network, the vibration wave model is used to determine the large-scale spatio-temporal change trend of the attack source, and the first-order autoregressive model is used to determine the small-scale spatio-temporal change trend. On this basis, the state equation of the power communication network attack source model is derived by using stirling interpolation formula, and the power communication network attack sources are aggregated. The two player attack defense game model is used to calculate the utility of network attack and network defense strategy, judge the utility of attack and defense, and evaluate the security of power communication network; and determine the entropy change rate threshold of power communication network, calculate the network entropy change rate, relative entropy and network packet baseline probability distribution, and design the power communication network attack detection steps to detect network attacks. According to the convolution neural network structure, the network activation function is selected. Through the network forward and back propagation, the network data categories are divided, the power communication network topology is determined, the network attacker and the attacked node are selected, and the attack source is located. Experimental results show that the proposed method can effectively improve the location accuracy of attack sources.