Aiming at the problems of low accuracy, difficult time feature extraction and separation of feature extraction and detection process of traditional electricity data anomaly detection methods, a power load data anomaly detection method based on deep convolution embedded LSTM auto encoder (DCE-LAE) is adopted. In this method, the long-short term memory network is integrated into the self-encoder architecture, the nonlinear feature extraction ability of the encoder and the timing feature memory ability of the long-short term memory network are used to improve the timing reconstruction accuracy of power load, and the deep convolution layer is embedded into the architecture to improve the receptive field to extract more time series features. In addition, the combination of convolution loss and reconstruction loss is used as the loss function for joint optimization to prevent the distortion of convolution embedding fine-tuning to the reconstruction space and further improve the reliability of the results. By comparing with other methods, the example simulation is proved that the anomaly detection accuracy and timing reconstruction ability of this algorithm are better than other algorithms.