In response to the current situation where electricity theft detection only uses a single electricity load, it is difficult to capture complex electricity theft characteristics, which leads to misjudgment in electricity theft detection, resulting in high false detection rate and low accuracy, this paper proposes a new electric theft detection method that integrates electric load, ambient temperature, time and phase line loss of corresponding station area. This article first builds MF-VAE（Variational Autoencoder for Multi-dimensional Feature extraction ）to extract the multi -dimensional characteristics of the user's electricity behavior. Then, based on TCN- Attention （Temporal Convolutional Network -Attention，TCN- Attention）,establish a discriminator model, to obtain the timing relationship of multi -dimensional electric stolen behavior through expansion convolution and causal convolution.At the same time, the convolutional attention module is introduced to assign the attention weight of each dimension feature to improve the performance and generalization ability of the model. Finally, Softmax classifier is used to accurately identify potential power stealing behavior in multi-source data. The experimental results show that the characteristics of electric theft behavior extracted by the proposed method are more abundant and diversified, which can effectively reduce the false detection rate of electric theft detection and improve the discrimination accuracy of electric theft behavior.