Generalization Ability Of Power Metering Anomaly Diagnosis Models

With the gradual increase of social demand for electricity,smart grid is the inevitable direction of the development of China’s power grid.As the core facility of power distribution and voltage conversion,substation is the key component of power grid.Therefore,strengthening the construction of smart substation is a necessary link in the construction of smart grid.Smart substation uses intelligent electronic equipment for information collection,measurement,monitoring,control and other operations,and further realize real-time monitoring,intelligent early warning,collaborative scheduling,decision-making assistance and other functions.Among them,power metering data analysis is becoming an important means of remote monitoring and management of substation operation status.In order to ensure the quality of power metering data,the anomaly diagnosis of power metering data in substation is of great significance.The anomaly diagnosis of power metering data in substation faces problems such as the small number of cases,the uneven distribution of samples,and the lack of "unknown" abnormal types of cases.This makes it difficult for traditional machine learning or even deep learning methods to work.In view of the above challenges,the existing power metering anomaly diagnosis models generally have the problem of insufficient generalization ability.With the upgrading of metering devices and the passage of time,the performance of models gradually degenerates and the recognition accuracy decreases.To sum up,the generalization problem of power metering anomaly diagnosis model can be summarized into two points: 1)In the face of small sample dataset model is easy to overfit and the generalization ability is insufficient.2)it lacks the ability to detect unknown anomalies.In view of the above generalization problems,this paper takes the metering data of substation as the research object and carries out research from the two aspects: Firstly,a power metering anomaly diagnosis model based on few-shot learning is designed for the small sample dataset.Based on the contrastive learning mode,a decision method based on the adaptive confidence coefficient is proposed,and the confidence interval threshold is integrated into the contrast loss function,so that the model can detect the "unknown" type anomaly better on the basis of considering the recognition performance of known anomaly.Secondly,aiming at the insufficient generalization ability of the model,this paper proposes an Auto-Ensemble method based on adaptive cyclic learning rate,which allows the automatic collection of multiple models with both accuracy and diversity in once training.By simply ensembling,the classification accuracy of the model can be improved,then the generalization ability of the model is improved.Through the above research,the contribution of this paper can be summarized into the following three points: 1)The problem of rare cases of metering abnormal data is solved.2)A method to detect the power metering "unknown" type anomaly is proposed.3)A method to improve the generalization ability of the power metering anomaly diagnosis model is proposed.In this paper,through a large number of experiments on electric dataset,it is proved that the proposed method can significantly improve the recognition ability of unknown anomaly of the power metering anomaly diagnosis model,the classification accuracy of the novel class reaches 89%.After ensembling,the classification accuracy of the novel class is improved by 0.1%,and that of the base class is improved by 1%.The generalization ability of the model is improved.To further verify the versatility of the proposed method,we conducted many comparison and simulation experiments on CIFAR and Omniglot datasets.The results show that the proposed method in this paper has achieved a larger performance improvement than the baseline model.Compared with the electric data experiment,the generalization ability has been greatly improved.The 20-way 1-shot Auto-Ensemble accuracy of the Omniglot dataset is increased by 2.85% compared to a single model.It is fully proved that the proposed method is suitable for traditional supervised learning and few-shot learning tasks,and can effectively improve the generalization ability of the model.