Abstract: Electricity theft not only disrupts the normal order of power consumption but also affects the quality of the power supply and the safe operation of the power grid. To solve the problem of diversification between normal power consumption and theft behavior of customers faced in electricity theft detection work, this paper proposes a method for detecting electricity theft in low-voltage stations based on multi-stage recursive data analysis. The first stage of the method identifies the suspected electricity theft area. A three-step analysis method based on the comprehensive fluctuation rate of line loss in the station area, the total-sub meters' current variance rate and the degree of time-overlap of sudden change points in the line loss and current curves are proposed for situations where the line loss is not significantly surging on that day, providing good conditions for the detection of suspected customers of electricity theft. The second stage proposes a time series similarity measure based on the most optimal set of special features. Based on the Euclidean distance measure of the numerical characteristics between curves and the dynamic time warping (DTW) algorithm measure of the morphological characteristics between curves, preliminary screening of suspected customers for electricity theft is achieved. The third stage proposes a support vector machine model for second-depth detection with optimized kernel functions and penalty parameters (OKPSVM), where the penalty parameters are optimized using an improved particle swarm algorithm (IPSO). The overall optimized support vector machine model (IPSO-OKPSVM) can improve the accuracy and applicability of deep power theft detection through arithmetic simulation and practical engineering applications.