Abstract: The recovery of stolen electricity is the ultimate goal of electricity theft detection, and accurate identification of electricity theft time is an important basis for accurately estimating stolen electricity. However, existing electricity theft detection methods focus on identifying electricity theft behavior and need more in-depth analysis of electricity theft time. There is an urgent need to study electricity theft time identification models based on electricity thieves' metering data to provide a basis for estimating electricity theft. Aiming to address the problem of electricity theft time identification, a semi-supervised electricity theft data classification method based on a transformer and one-class support vector machine (OCSVM) is proposed. Firstly, the user load data is divided by day, and the problem of identifying the time of electricity theft is transformed into the problem of power theft daily load data discrimination. Secondly, the Transformer is used as the reconstruction model to learn the user's normal power consumption patterns and regularity to reconstruct the reconstructed value based on the user's daily load data. Finally, the reconstructed error curve is used as the input of the OCSVM to construct the decision boundary of normal power consumption behavior. Then the electricity theft data is identified to realize the electricity theft time identification. An example analysis was carried out based on smart meter user data in a southern province to verify the feasibility and effectiveness of this method. The experimental results show that this method has good sensitivity and robustness.