Research On Multi-classification Method For Smart Meters Fault With Multiple Distribution Patterns

As an important terminal of smart grid,smart meters play an important role in supporting the normal consumption of users and the stable operation of the power system.Smart meters not only have the traditional functions such as data metering and display,but also possess the extended functions of remote reading,bidirectional communication and so on.With the increasing number and functions of smart meters,there are diverse and complex smart meter faults.It is difficult for maintenance personnel to judge the fault types by their own experience,resulting in problems such as delayed or improper treatment for meter faults.Therefore,accurate prediction of smart meter faults is of great guiding significance for making reasonable maintenance plans.Since the fault type is affected by many factors,it is an effective way to construct the mapping relationship between the multi-dimensional attributes of smart meters and their fault types by using machine learning method.However,there are different occurrence frequency and class overlaps of different smart meter fault types,which greatly increases the prediction difficulty of smart meter fault.Focusing on the multiclassification task of smart meter faults,aiming at the characteristics that there are multiple patterns of instances with the same fault type and class overlaps with different fault types,imbalanced binary classification is studied based on one-to-rest classification framework in this paper.The main work is as follows:Firstly,a smart meter fault classification method based on generated distribution optimization is studied.Most existing methods only balance the number of samples,which may aggravate overlaps and not fully consider the data distribution in the overlapping area,leading to performance degradation.In this paper,a sample distribution optimization method based on multi-objective evolutionary algorithm is proposed.This method transforms the problem of sample rebalancing of different classes into the problem of data distribution optimization.The Euclidean distance is used to calculate the nearest neighbor sample of the selected sample,and the verification sample is generated by linear interpolation between the selected sample and its neighbor.Using the classification performance of the overlapping dataset and the constructed verification dataset as the optimization objective function,the data distribution optimization model of the generated samples is established to ensure that the generated samples can effectively assist the classification model in constructing a more accurate decision boundary.The differential variation strategy is applied to obtain the generated samples,which increases the diversity of the generated samples.Meanwhile,data complexity is introduced to modify the original NSGA-II so that datasets with lower data complexity are retained to avoid aggravating overlaps.Secondly,the classification method of smart meters faults for overlapping and within-class imbalance is studied.In order to solve the problem that the classification performance is limited by the ability of the classifier to fit the complex decision boundary,this paper propose an imbalanced binary classification method via space mapping using normalizing flows with class discrepancy constraints.The flow-based model is used to ensure the consistency of original and latent distribution by using the accuracy and reversibility of the flow-based model.The latent space with higher linear separability is obtained by introducing class discrepancy constraints on the latent distribution.Specifically,the global constraint maps sub-clusters of the same class to the same latent area and sub-clusters of different classes to the different latent area to alleviate within-class imbalance.The local constraint increases the interval of different classes to alleviate and reduce the class overlaps.Finally,the classification method for smart meters faults considering neighbor information is studied.Existing methods for imbalanced classification usually construct a mapping between individual instance features and its corresponding class label.As instances of different classes in the overlapping area have similar features,it is difficult for the classifier to accurately classify instances based on individual instance features.A smart meter fault classification method based on multigrain neighbor graph is proposed.Firstly,instance in the original dataset is selected as target instance(TS).The nearest neighbor graph(NNG)is constructed by taking the TS and its nearest neighbor instances as nodes and the line between the TS and its nearest neighbor instances as edges.According to the number of selected nearest neighbor instances,multigrain neighbor graph(MNG)is constructed to realize information expansion of TS and the number expansion of training set.An encoder is constructed to mine node features,and the graph attention is used to adaptively aggregate the information of the nearest neighbor instances to the TS based on the embedding node features and adjacency matrix,so as to realize the effective mining of similar sample differences.For a given test instance,more accurate and robust results are obtained by integrating the classification results of the MNG of test instance.