In order to improve the quality of power monitoring and avoid interference from abnormal data on monitoring results, a deep learning based automatic capture method for abnormal data in power monitoring networks is proposed. Firstly, the gradually aligned fusion network in deep learning is used to extract the feature sequence of power monitoring network data, and the feature fusion of power monitoring data is completed through the calculation of weights and feature channels. Finally, input the fused data features of the power monitoring network into the support vector machine, and complete the classification of abnormal data and normal data by constructing the optimal hyperplane, achieving automatic capture of abnormal data. The experimental results show that compared with existing anomaly data capture methods, our method has higher recall and shorter capture time.