To help reduce the false alarm rate of electricity theft detection a method of detection based on Bayesian optimization and an improved XGBoost model is proposed. First, a detection index of electric theft is constructed based on the data of electric energy data acquisition and marketing systems, and then the XGBoost model is used as a classifier, and the Focal Loss and increased classification threshold are used to optimize the mode. The MCC of the validation set is used as the objective function, and the optimal Focal Loss parameters and classification threshold are found using Bayesian optimization to adjust the parameters, which further reduces the false alarm rate. The detection method misclassification rate is reduced further. Numerical simulation is carried out based on actual power user data,. The results show that the proposed method has higher accuracy than Adaboost, a BP neural network and SVM. This work is supported by the Science Research Key Project of Hunan Education Department (No. 19A349) and the Science Research Outstanding Youth Project of Hunan Education Department (No. 19B396).