In view of the influence of industrial users′ industry attributes on their power consumption patterns, a power consumption anomaly identification method considering industry relevance is proposed in this article. Based on the real industrial consumer power consumption data, the typical load characteristic curves of each industry are generated, and the improved grey correlation degree algorithm is used to calculate the relevance between the power consumption characteristics of power users and the typical power consumption characteristics of the industry. In this way, the industry relevance characteristics of users are achieved. The multi-head attention (MHA) is used to extract the features contained in load sequences. Combined with the industry relevance features, the reconstruction error provided by the variational autoencoder (VAE) is used as the anomaly decision metric to formulate the MHA-VAE depth anomaly detection model to identify various types of industrial users′ power consumption anomalies. Results show that, the accuracy, detection rate and false detection rate after introducing users′ industry relevance are 96. 84% , 98. 02% , and 4. 35% , respectively. Compared with only considering the load characteristics of users, the accuracy is increased by 1. 06% and the error detection rate is reduced by 2. 24% .