Prediction Of Power User’s Outage Sensitivity Based On XGBoost Algorithm

With the continuous development of smart grid technologies such as electricity consumption information collection systems and new generation distribution automation systems and Internet+power marketing,power companies have more and more ways to obtain user data,and the amount of power data they obtain is also increasing.In this era of big data with the Internet of Everything,how to effectively extract effective information from massive data has become one of the research directions of many researchers.The application of user profile technology to the power industry,combined with big data technologies such as data mining and machine learning,can accurately predict the sensitivity of power users to power outages,which is conducive to giving full play to the power supply company side power outages,load information and user side payment of electricity bills,complaints and other data inherent Value;it is conducive to the power companies to give full play to their subjective service initiative,to respond to service needs in a timely and targeted manner,and to support the precise marketing and differentiated services of power users;it is more conducive to the power companies to be reasonable in the new competition of power system reform Allocate resources,enhance market competitiveness,and grasp first-mover advantage.Therefore,this article focuses on the prediction of power users’ outage susceptibility from the following three aspects:(1)Generate power users’ outage sensitivity labels.In the actual application of electric power,although electric power companies have abundant marketing data,the complexity of the data makes it difficult to obtain power outage sensitivity labels directly from power data,and the existing power outage sensitivity labels are only divided into two types:sensitive and non-sensitive.Class,which is far from meeting actual business needs.In response to these problems,this article starts from the types of power outages,introduces binary codes to indicate power users’ power outage sensitive categories,and further subdivides the power outage sensitive labels.Specifically,based on actual business needs,this article first builds a power user outage sensitivity classification label system,then converts the label system into a hierarchical structure and uses the analytic hierarchy process to assign weight to the outage sensitivity label,and finally normalizes it through indicator data,Power outage sensitivity conversion and other operations have generated a power outage sensitive label represented by a binary code.This provides a scientific and practical label classification result for the subsequent prediction of power users’ power outage sensitivity.(2)Processing of the data set of power outage sensitive user prediction under the multi-source feature system.In order to ensure the quality of the data used to predict the power outage sensitivity of power users,this article mainly processes the collected actual data from multiple perspectives such as data cleaning,feature extraction,and unbalanced data processing to improve the accuracy of the prediction model.And stability.For different types of data,this paper adopts a variety of different feature extraction methods,and constructs a multi-feature system of the power outage sensitivity prediction model.At the same time,in order to solve the problem of uneven distribution of actual data categories,this paper proposes a sampling method based on SMOTE+multiple replacement RUS.Through the performance comparison of the data set before and after sampling on the training of multiple classification algorithms,the sampling is verified The effective practicability of the method provides strong support for the excellent performance of the power user’s power outage sensitivity prediction model.(3)Construct a power outage sensitivity prediction model based on the improved PSO-XGBoost algorithm.In order to accurately predict the power user groups that are sensitive to power outages,this paper considers the impact of feature selection on the performance of the model,and proposes an improved algorithm based on PSO-XGBoost.Specifically,the algorithm uses the number of feature sets sorted based on the XGBoost algorithm as the parameter to be optimized into the PSO algorithm,and then uses the PSO algorithm to optimize the hyperparameters of the XGBoost algorithm.In order to verify the practicability and efficiency of the improved algorithm,this paper designs two sets of performance comparison experiments.The first is the performance comparison between the GS-XGBoost algorithm and the improved PSO-XGBoost algorithm.The second is the performance comparison between the logistic regression algorithm,decision tree,GBDT,random forest,AdaBoost and other common classification algorithms and the improved PSO-XGBoost algorithm.The experimental results show that the model used in this paper shows better prediction accuracy and stronger generalization ability than other models in the prediction of power user outage sensitivity.This is of great significance for the prediction of power outage sensitivity of actual power users.