Non technical losses in power systems, represented by abnormal electricity consumption by power users, will typically result in significant increase in the operating costs of power supply companies. Firstly, deep neural detection method for abnormal electricity consumption by power users is proposed. Based on the characteristics of electricity consumption by power users, a deep confidence network（DBN）is used to extract features from the original electricity load data and obtain corresponding features.Then, feature classification is completed using an extreme learning machine（ELM）, thus establishing a basic model for detecting abnormal electricity consumption by power users. Finally, an improved fruit fly optimization algorithm（IFOA）to optimize the network weights and inter layer bias parameters of DBN is proposed,thereby obtaining an abnormal electricity consumption detection model for power users based on IFOA-DBN-ELM. Experimental results show that compared with other detection methods, the accuracy, precision, and detection rate of the method proposed are significantly higher, and false detection rate are lower than other methods. It can accurately detect power users with abnormal electricity consumption behavior and help reduce the operating costs of power supply companies.