To effectively ensure the stable operation of the power system and identify and prevent potential faults in a timely manner, this study proposes an anomaly detection method for power dynamic data combining deep learning and a multivariate data drift detection algorithm. Firstly, Long Short-Term Memory (LSTM) is chosen as the deep learning model in dealing with time series data. The LSTM model can learn the normal behavior pattern of the power dynamic data by extracting the depth features, and identify the abnormal behavior through the prediction error when the data is abnormal. To further improve the accuracy and efficiency of anomaly detection, a multivariate data drift detection algorithm is introduced. By comparing the statistical differences between the current data window and the historical data window, the algorithm sensitively captures the changes of data distribution and finds potential anomalies in time. In the experiment, the accuracy, recall, precision and F1 score of detection are significantly improved. The proposed anomaly detection framework not only makes full use of the advantages of deep learning in feature extraction and modeling but also combines the sensitivity of a multivariate data drift detection algorithm in data dynamic change monitoring, providing a more comprehensive and accurate method for anomaly detection of power dynamic data.