Abstract: Microphone array can detect partial discharge (PD) of power equipment in a non-contact and flexible way. However, existing methods lack the consideration of data characteristics of acoustic array, and the researches on identification of PD type are insufficient. Considering the correlation and imbalanced distribution features, this paper firstly analyzes the temporal and spatial correlation characteristics of microphone array data. Secondly, based on one-dimensional convolutional neural network and "squeeze-and-excitation" correlation extraction method, a PD pattern recognition model based on spatial and temporal correlation feature fusion strategy is proposed. Finally, the loss function adjustment method and data distribution adjustment method are used to deal with the imbalance between different PD classes. Simulations show that, compared with the methods in which the correlations are not taken into consideration, the methods proposed in this paper enhance both the precision and recall by more than 12%. Compared with the methods in which the data imbalance is not taken into consideration, the methods improve the precision and recall by over 60%, respectively. These results affirm the essential need to consider both correlation and imbalance characteristics in acoustic array based PD recognition.