Non-technical loss in power grid not only has a significant impact on the economic benefits of the power company, but also poses a serious threat to power quality and operational safety of the power system. In addition, measures taken by malicious users to seek profits grow in complexity, resulting in traditional detection methods gradually falling to limitation. Implementation means for non-technical loss based on electricity consumption curve are studied and tampering strategies used to generate false data are summarized. Behavior features of power users are extracted from the electricity consumption curve and associated with the results of electrical tampering implementation by bidirectional long short-term memory network. Finally, a multi-level neural network architecture is designed and deep learning is utilized to solve the multiclass classification problem of the feature sequences. Simulation based on actual power consumption dataset of a certain area shows that the research content can realize an effective detection of non-technical loss as well as identification of specific tampering strategies.