Given the problems of a complex identification algorithm and the low recognition rate of a power quality disturbance signal, a new method of applying a long-short term memory neural network to power quality disturbance signal recognition and classification is proposed. First, a long-short term memory neural network is built in Tensorflow to establish a power quality disturbance signal classification model. Secondly, the classification model is used to supervise the original data of the power quality disturbance signal, and the deep features of the disturbance signal are extracted and connected. The Softmax classifier outputs the recognition rate of each disturbance signal. Finally, a two-dimensional trajectory image generated by the power quality disturbance signal through the recursive map is used as the input of the classification model, and the disturbance model is classified by the training model. The simulation results show that the classification model has good classification accuracy for the one-dimensional and two-dimensional representation of power quality disturbance signals, and can effectively identify seven single disturbances and six composite disturbance signals. This work is supported by National Natural Science Foundation of China-Shanxi Coal-based Low-carbon Joint Foundation (No. U1610120).