Abstract: With the rapid development of smart grid in China, the monitoring data of power grid presents the trend of diversification, high speed and quantification. In order to fully exploit the potential value of power big data, and realize the automatic recognition and location of abnormal regions in power grids, a method of locating abnormal events was studied based on random matrix theory (RMT) and convolutional neural networks (CNN) in this paper. Firstly, the power grid was divided into several subsystems according to the network relationship, and a monitoring matrix was constructed for every sub-area. Then, taking RMT as the feature extraction method, the extracted feature vector of partition matrix was used as input for the monitoring system. And a CNN model was built according to the characteristics of grid monitoring data and abnormal recognition requirements. Finally, a dataset was constructed based on partition matrix feature vector to obtain an effective CNN model for automatic location of abnormal events. Taking the three-phase short-circuit fault of IEEE39-node power grid model as an example, the analysis results show that the preprocessing method of extracting feature vector by RMT can effectively reduce the data dimension and improve the fault location accuracy of CNN model. The partitioned RMT-CNN model can effectively locate the location of abnormal events in the grid with a location accuracy of 97.96% and precision of 98.65%.