Research On Business Anomaly Detection And Fault Tracing Technology Of Smart Grid Dispatch Control System

The Smart Grid Dispatching Control System(D5000 System)realizes real-time monitoring,online stability analysis,dispatching business management and other functions in power grid operation,which improves the dispatching organization’s ability to control large power grid,optimize the allocation of resources in a large range,and deal with major power grid failures.As a typical information physical system,the normal data obtained by the state monitoring system is far more than the abnormal data in a large number of real-time monitoring data,and the logical relations of its components are complex,varied and change dynamically with the adjustment of business functions,so it is difficult to obtain the logical topological relations.Once the system business abnormal and fault,fault propagation range is wide and fast will cause a series of components of the alarm,the stable operation of the power grid has a great impact.With the continuous upgrading of system services and the continuous expansion of network,the traditional anomaly detection method of threshold setting based on expert experience fails to consider the correlation and mutual influence of data dimensions,which leads to a high detection error alarm rate and is difficult to find some deep-level potential system business anomalies,which cannot meet the requirements of high precision system anomaly detection.At the same time,when it is difficult to obtain the logical topological relationship of the system,the operation and maintenance personnel cannot carry out fault source quickly and accurately.Based on the ideas of machine learning and information theory,this paper studies the business anomaly detection and fault traceability technology methods of smart grid dispatching and control system.The main work is as follows:(1)The anomaly detection method for dispatching control system under the condition of data imbalance is studied.Aiming at the problem of anomaly detection in D5000 system,an ensemble imbalanced classification method based on model dynamic selection driven by data partition hybrid sampling is proposed.First,a data partition hybrid sampling(DPHS)method is proposed to balance datasets.In particular,the data space is divided into four regions according to the majority class proportion in minority class neighborhoods.Second,we present a boundary minority class weighted over-sampling(BMW-SMOTE)method where the weight of each minority class instance is calculated by the ratio between the majority class proportion in the neighborhood of the current instance and the sum of all these proportions.calculates the number of compounds in each minority class neighborhood,and carries out oversampling in the way of random linear interpolation.Finally,we present a model dynamic selection(MDS)strategy.Three ensemble learning models are built:the original model biased towards the majority class,the local region strengthen and weaken model,and the mixed model biased towards the outer boundary.Among them the local region strengthen and weaken model adopts the available balanced dataset.The model for each test instance is selected adaptively according to the imbalance degree of its neighbors.The effectiveness of the proposed method is analyzed by comparing the typical method with the public dataset and the D5000 system business dataset.(2)The fault tracing method for dispatching control system based on data-driven is studied.Aiming at the difficulty of fault source when the topological relationship between system components is unknown,a fault traceability method for power dispatching control system based on information difference graph model is proposed.First,aiming at the problem that the information measurement cannot be directly obtained for D5000 system,the feature interval clustering mean discretization method is proposed.The clustering center of each data feature time series collected is obtained through the k-means algorithm,which is taken as the endpoint dividing the discrete interval and the interval mean is calculated as the discretization result.Second,the information difference matrix is established according to the information measurement rate before and after the alarm.The matrix elements are composed of the information entropy of the system components and the normalized difference ratio of the transfer entropy between the components respectively.Finally,an information difference graph model is established according to the characteristics with high variation alarm information and the interaction information among them.The state and link relationship of nodes in the graph model are defined,the mapping table between nodes and their weights and fault degree is constructed,the components in the table with the highest fault degree are marked,and the interaction relationship between components in the graph model is traced for fault source.Experimental results on simulation data,open-source distributed system dataset and D5000 system dataset show that the proposed method has significant advantages over the canonical correlation method in terms of accuracy,recall rate and other indicators.(3)The fault propagation path reasoning method for dispatching control system based on sorting learning is studied.Aiming at the problem that the fault correlation propagation leads to low troubleshooting efficiency and the system cannot directly backtrack the fault propagation path from the time series,a fault sorting learning method based on ELFC and LambdaMart algorithm is proposed.Firstly,the dynamic process of D5000 system logs is decomposed into normal segment,abnormal segment and recovery segment,and the time series of abnormal segment and recovery segment are intercepted to obtain the sequence set queue advancing according to time by feature selection.Second,input sequence set generated by ELFC queue algorithm initialization LambdaMART model,constant iterative learning system in the model correlation function to obtain the minimum loss,use log of word frequency,inverse frequency components and the component related log length to fit sequencing scores,the final training completed ELambdaMART model to study sort of test set,get the failure of objective mining association recommended list,using differential information graph model,filtering and pruning the independent components,fault path tracing.Experiments on public datasets and D5000 system dataset show that the proposed method has improvements in NDCG,MAP and shortening training time,which verifies the effectiveness of fault propagation path inference determination for a D5000 system.