Existing detection methods of electricity theft have low detection accuracy for new means of electricity theft. This paper proposes a new theft detection method based on multi-level non-negative sparse coding and SVM. Using the monthly electricity consumption curve as the detection object, firstly, the multi-level electricity consumption pattern characteristics of the sample are extracted based on the multi-level non-negative sparse coding; next, the numerical statistical characteristics of the sample are extracted based on the electricity theft scenario analysis; then, the fusion detection features of the above two characteristics are input into the SVM classifier for electricity theft detection. Finally, the Irish smart meter data set is used as the example to verify the effectiveness of the proposed method, showing the improved accuracy and recall rate of the detection.