Research On Business Anomaly Detection Method Of Power Dispatching Automation System Based On Machine Learning

The Smart Grid Dispatching Control System(D5000 system for short)improves the ability of multi-level dispatching to jointly handle major power grid accidents and plays an important role in maintaining the safe and stable operation of the power grid.Once the system businesses run abnormally,the power grid will be greatly affected or even paralyzed,which leads to huge economic losses.Since there are many characteristics in the D5000 system,such as an amount of business services,complex business-level associations,large business data dimensions and many types of data,when dispatchers check whether the system businesses are running abnormally,the current method,setting a threshold for a single state quantity,does not take into account the relationship between the indicators,and cannot adapt to the dynamic changes in businesses,which makes the anomaly detection accuracy low.Anomaly detection methods based on machine learning provide some solutions to solve the above problems,but the existing methods have some disadvantages,such as low detection efficiency and low accuracy.In order to make the system more intelligent,help dispatchers understand the system’s business operation status in time,and make the businesses run stably and normally,based on the machine learning,this paper conducts research on some key technologies,such as the threshold setting for anomaly detection model,anomaly detection on static data and anomaly detection on real-time streaming data.The main research works are as follows:1)In order to realize the intelligent setting of the threshold,and avoid adjusting the anomaly ratio in the algorithm model repeatedly and manually,an adaptive threshold setting method based on the golden section rate is proposed.Firstly,based on the application background of the system,the requirements of the system’s business threshold setting are analyzed.Secondly,based on the machine learning algorithm and the sorted anomaly scores,a function can be obtained by means of polynomial fitting.After using the golden ratio for analyzing the function,this paper explores the distribution of the data and selects a more appropriate value as the model threshold.Finally,this paper uses this adaptive threshold setting method to improve the isolated forest algorithm.In order to verify the feasibility of the adaptive threshold setting method in the field of anomaly detection and make the preparation for subsequent research on a new unsupervised anomaly detection method,this paper analyzes the performance changes between the original isolated forest algorithm and the improved one,using various public datasets and actual business data of the D5000 system.2)Aiming at some disadvantages of the current anomaly detection algorithms,such as low detection efficiency and low anomaly detection accuracy for the data with high-dimensional or numerical dimension and logical dimension coexisting,an unsupervised anomaly detection method using center offset measurement based on leverage principle is proposed.Firstly,this paper analyzes the requirements of business anomaly detection and the disadvantages of applying existing methods to this system.Secondly,based on machine learning and leverage principle,the abnormality of the data to be measured is evaluated by measuring the offset of the center of the original dataset.This paper also uses the skewness to obtain the spatial distribution status of the dataset and formulates the corresponding abnormality judgment adjustment strategy.Combining with the adaptive threshold setting method,a multiplication leverage anomaly detection model can be constructed.Finally,the proposed method is compared with various typical anomaly detection methods,and the advantages of this method in business anomaly detection are analyzed,using a number of public datasets and D5000 system business datasets.3)In order to improve the accuracy of streaming data anomaly detection,an anomaly detection method for streaming data based on roulette is proposed.In addition,aiming at the problem that the current updating strategies of model cannot compare the performance with each other,an evaluation criterion for the updating strategy of streaming data anomaly detection model is proposed.Firstly,based on the situation that businesses would change when the system is running,this paper analyzes the requirements of system’s online anomaly detection.Secondly,based on the multiplication leverage algorithm and the roulette strategy,this paper stores data in the cache area selectively and uses the data in the cache for updating the multiplication leverage model.Finally,by analyzing the performance of different streaming data anomaly detection algorithms,and the performance fluctuation between the original model and the improved,a streaming data anomaly detection evaluation criterion is proposed.In the experiment,it is also used as one of the evaluation criteria to compare this method with other existing streaming data anomaly detection methods.