There are many classification-based methods for abnormal electricity consumption detection now,but most of them are based on short-term electricity consumption to judge long-term electricity consumption behavior. It is difficult to determine the threshold and ratio of these methods. In engineering application,the distribution of power consumption data in different regions and time periods is quite different,so the proportion and threshold value are quite different. It is difficult to apply the fixed proportion to all user data. To solve this problem,a method for judging abnormal electricity consumption based on reinforcement learning is proposed,which innovatively uses reinforcement learning model to dynamically generate threshold for different data sets. Firstly,the abnormal probability of the short-term behavior of several users output by the classifier is obtained. Then,the dynamic threshold is obtained by inputing the probability into the deep recurrent Q network (DRQN)of the enhanced learning model,where,the dynamic threshold can be Judgment threshold and judgment ratio as well. The experimental results show that,compared with the traditional voting method of manual parameter adjustment,this method has a significant improvement in the evaluation index,and also has a good performance in data sets with large differences in data distribution. It shows that this method has strong generalization ability in the real environment with complex data types.