Research On The Classification Of Power Outage-sensitive Users Fused With Multiple Behavior Characteristics

With the construction of power grid infrastructure and the vigorous popularization of electrical equipment,electricity plays an increasingly important role in people’s lives.According to the statistics of the 95598 business department of a city in the south,complaints about power outages and power supply quality accounted for 52.6% of the total number of complaint work orders in the past year.The problem of power grid outages has attracted increasing attention from customers.Predicting whether a user belongs to a power outage sensitive user in advance is not only helpful for grid companies to provide timely and accurate services for users in a power outage accident,to improve users’ power consumption experience,but also to reduce the number of 95598 consultations and customer complaints.Based on the data of blackout users collected by multiple business departments of the power grid company,this paper conducts the following research on the classification of blackout sensitive users: 1.Missing values and abnormalities of data on blackout users that contain both structured and unstructured data Value and text data segmentation and other preprocessing work,extract the user’s power consumption behavior,call behavior and feedback behavior related features to build a blackout user label system,and effectively build multi-source heterogeneous blackout user data into a complete and unified feature Vector space;2.Due to the high dimension of the primary selection feature set and containing many different types of features,different feature selection methods are proposed for different types of features,and a highdimensional feature selection method based on dual channels is proposed.By improving XGBoost The algorithm’s numerical feature selection method screens out the optimal numerical feature subset,and effectively reduces the dimensionality of the text feature by applying the improved firefly algorithm to the text feature screening process;3.Proposes an undersampling method based on sample weights for power outage users Effectively process the unbalanced data,and extract corresponding features in turn according to the preferred feature set of the blackout user to characterize each data sample,use it as an input sample for model training to establish a blackout sensitive user classification model,and judge this based on new user data.Whether the user belongs to a power failure sensitive user,after experimental verification,the accuracy rate of the classification model proposed in this paper on the high-dimensional unbalanced data set reaches 90.38%,and the recall rate is 92.23%.