Traditional electricity theft detection methods relying solely on single power consumption loads and struggle to capture complex theft patterns, resulting in high false detection rate and low accuracy. Thus this paper proposes a new electric theft detection method that integrates electric load, ambient temperature, time and corresponding phase line loss of the station area. First, a multi-dimensional feature extraction variational autoencoder (MF-VAE) is constructed to extract the multi-dimensional characteristics of the user’s electricity consumption behavior. Then, an attention temporal convolutional network (ATCN) is employed to establish a discriminator model, where expansion convolution and causal convolution are used to obtain the temporal relationships of multi-dimensional electric theft features. Meanwhile, a convolutional attention module is introduced to assign the attention weight of each dimension feature to improve the performance and generalizability of the model. Finally, a Softmax classifier is used to accurately identify potential power theft behaviors in multi-source data. Experimental results show that the characteristics of electric theft behaviors extracted by the proposed method are more abundant and diversified, which can effectively reduce false detection rates and improve the accuracy of electric theft behavior identification.