Abstract: In recent years, electricity theft detection methods for high-loss lines have been applied in a large area of engineering,which plays an important role in reducing the false alarm rate of electricity theft detection and promoting the engineering application of data-driven electricity theft detection. However, it is still a difficult problem to accurately detect the special transformer users of non-high-loss lines. Based on the characteristic that some users who steal electricity have abnormal power consumption peaks in the practical experience, a method of identifying users with electricity theft actions based on long short-term memory（LSTM）autoencoder（AE）with load peak features is proposed. First, the peak characteristics of power consumption that distinguish normal users and users with electricity theft actions are extracted by analyzing the curves of typical users with electricity theft actions.Then, the LSTM-AE model is constructed to reconstruct the input and obtain the fitting value by combining this feature and the periodic law of user time-sharing data. The adaptive threshold is set based on the mean square error between the fitting value and the real value, so as to identify the suspected users of electricity theft and provide specific warning peak time. Finally, an example is used to analyze the actual power consumption data of special transformer users. The results show that the proposed method is superior to the comparison method in terms of accuracy, hit rate and false alarm rate.