Research On Ensemble Classification Method For Smart Electricity Meters Fault

With the improvement of the intellectualization degree of the power industry,mart electricity meters have an increasingly wide coverage and have been gradually replacing the traditional non-intelligent electricity meters,due to their advantages in information collection and storage,real-time monitoring and control,network communication and information interaction,and other performance advantages.At the same time,because of the significant increase in the complexity of function and structure,the failure rate of smart electricity meters in the actual complex working conditions was also increased.In order to improve the safety and economy of electricity consumption,it is important to identify and predict the faults of electricity meters reliably.On the basis of the historical fault data of smart electricity meters,this thesis uses statistical principles and machine learning methods to analyze the attribute characteristics of electricity meters and studies the fault classification and prediction of electricity meters.The main work is as follows:Firstly,the meaning and statistical characteristics of the attribute of fault data of smart electricity meters are analyzed in detail,respectively fr-om the aspect of categorical attribute and numerical attribute.Classification attributes are filtered by correlation.And the attribute field was preprocessed.It provided a reliable basis for subsequent feature engineering and model building.Secondly,this thesis makes a comparative analysis of the multi?category classification methods based on supervised learning.At the same time,the performances of various classifiers on fault data sets are also compared and analyzed.Thirdly,in this thesis,a probability-based categorical feature extraction scheme is designed in view of the performance of smart electricity meters in decision tree and k-nearest neighbor classifier and the feature of large proportion of categorical attributes.This thesis applies the idea of ensemble learning to k-nearest neighbor classification algorithm and integrates several similarity measurement methods.The rules for fault classification of smart electricity meters are obtained with the improved performance of the classifier.Finally,in this thesis,k-nearest neighbor classifier is used to get the importance weight of each attribute,and a feature weighting based XGBoost classification method is designed.The performance of the XGBoost classification method is improved by changing the weight of the input data while extracting the available information of the classification features.