The human inspection process used in classical power theft detection is expensive and ineffective in obtaining tagged data. With the development of smart grids that store rich user electricity use data, machine learning-based methods for detecting power theft are emerging quickly. However, the question of how to increase power theft detection accuracy without depending on labeled data remains a challenging one to answer. Therefore, In order to detect power theft, this research suggests a strategy based on contrast predictive coding support vector data description(SVDD). To do this, it first uses contrast predictive coding to extract long-term pattern characteristics from user data that most accurately reflect the user's power consumption habits. And builds positive-negative sample pairs using gated recursive units for comparative learning. The SVDD classifier is then trained using the characteristics retrieved by contrast predictive coding to determine the center and radius of the related hypersphere. Finally, To determine who is stealing electricity, the relationship between the radius of the hypersphere and the distance between the center of the hypersphere and the electricity consumption data to be examined is employed. The algorithm is simulated using actual State Grid data, and the experimental findings demonstrate that, in comparison to the comparative approaches, the model presented in this study has a better classification performance and a reduced false alarm rate.