Research On Abnormal Detection And Classification Of Power System

Cyber attacks on the power system and emergency events such as network attacks will cause abnormal power outages and interruptions at the user end,causing huge economic losses and security incidents.Therefore,the method of power system anomaly detection and classification based on machine learning has important research value and practical significance.Existing supervised power system anomaly detection and classification methods cannot identify unknown events or attacks and their performance drops sharply when training data is insufficient,while unsupervised power system anomaly detection and classification methods have low performance and cannot provide specific information about anomalous events or attacks.Two different power system anomaly detection and classification methods are proposed to solve the above problems.The main contents and innovations of this thesis are as follows:(1)For the problems of low power system anomaly detection recall rate and performance degradation,this paper proposes a combined power system anomaly detection method that can automatically select the best detection algorithm based on the amount of input training data.The main innovation is that when the training data is sufficient,the enhanced multi-granularity cascade forest is used for anomaly detection,and the gradient random decision tree is used to replace the completely random forest in the cascade forest,which further enhances the integration and robustness of the cascade forest.The improved multi-granular cascade forest not only has the excellent characteristics of deep learning algorithms,but also effectively shortens the time required for training and reduces the complexity of the model.When the training data is insufficient,genetic algorithm is used for anomaly detection,and the genetic algorithm is improved by floating-point number coding,which speeds up the convergence speed,simplifies the operation difficulty,and shortens the training time.Anomaly recognizers based on genetic algorithms only need Normal events in less than 20% of the training data are enough to build an anomaly detector,effectively reducing the amount of training data,and also maintaining a high recall and accuracy rate.(2)For the problem that supervised power system anomaly classification method cannot identify unknown events or attacks and the complexity of deep learning algorithms is too high,while the unsupervised power system anomaly classification method has low performance and cannot provide specific information about abnormal events or attacks,this paper proposes two different methods for power system anomaly detection and classification.The main innovation is that when the training data is sufficient,an adaptive method based on an improved multi-granular cascade forest is used for power system anomaly detection and classification.The method uses semi-supervised learning and contour coefficients to design a system update module to complete the classifier automatic updates and evolutionary iterations.In addition,the use of contour coefficients can effectively identify new attacks and unknown events.When the training data is insufficient,the proposed multi-layer detection and classification method based on semisupervised clustering can also maintain a high attack recall and classification accuracy.The label expansion method effectively increases the amount of training data and solves the problem of classifier performance drop caused by insufficient data,and the idea of semi-supervised clustering is used to provide specific information on detected attacks or events.Finally,similarity measures are used to screen out unknown attacks or events after clustering.