Aiming at the problem of abnormal data in the application of smart meters, by defining any abnormal power consumption instance or trend beyond the normal use mode of each load, a data-driven model based on different types of characteristics such as load, context and environment is designed, and four different unsupervised models based on regression, neural network, clustering and projection are evaluated The performance of the learning method in the actual smart meter data anomaly detection is analyzed. The results show that different anomaly detection methods have different detection ability for different types of anomalies, and their performance depends on the feature set used to train the method. Therefore, for each anomaly detection method, different types of features should be carefully examined.