Machine Learning Based False Data Attack Detection In Power System

In recent years,there has been an increasing social demand for the reliability of power supply and the safety of power systems.State estimation is the most important part of the power grid dispatching control.However,due to the deep coupling and integration of communication networks and power systems,the measurement data required for state estimation of power system is extremely vulnerable to cyber attackers,making it difficult to ensure grid data accuracy.The vulnerabilities of communication networks pose a high threat to the safety and reliability of power systems,and network threats have a significant impact on the security of cyber physical power systems.Fales data attack is a new type of network attacks that meets the consistency of power grid.This can bypass traditional bad data detection mechanisms and maliciously tamper with grid data.This affects the results of state estimation and triggers faults,causing physical damages to the operating state of the system.Therefore,while promoting the informatization and intelligent construction of the power grid,focusing on the data security of the smart grid and ensuring the accuracy of the power grid data are currently urgent issues to be solved.Firstly,we introduce the general principle of power system state estimation and the mechanisms of traditional bad data detection.Then,we explain the attack mechanism of false data in linear and non-linear state estimation models respectively,and develop the corresponding detection methods to defend the malicious data attacks under various situations.Machine learning techniques have also been employed to improve the detection accuracy of static and dynamic false data attacks.For false data attacks under the DC state estimation model,the power characteristics of the data samples are first combined with the spatial distribution characteristics.We achieve the progress of feature extraction and dimensionality reduction of high-dimensional power data through SSAE.Then,the genetic algorithm is used to optimize the hyperparameters of the XGBoost classifier to help identify the anomaly data.Based on this,a semi-supervised FDA detection model based on SSAE-GA-XGBoost is proposed to realize the identification of abnormal power data.For malicious data attacks under a dynamic evaluation model,this paper takes a sample of a 3-month continuous power dataset provided by CAISO as example and uses EEMD decomposition method to fully explore the timing characteristics of the input signal under different frequency components.We identify the complexity of feature components by introducing the concept of fuzzy entropy.On this basis,the GRU neural network under the attention mechanism is used to optimize the weight coefficients of the feature components with strong decision-making properties.Through analyzing the residuals between the predicted value of the state quantity and the actual value obtained by AGRU,we judge whether the power grid is attacked by malicious data.In addition,by evaluating and analyzing the system state at the sampling time when the false data is injected,the recovery of the actual power data is realized by using the EEMD-AGRU based network model to predict the state at the time of the attack with the historical data samples to ensure the safety and integrity of power data and improve the resistance of the smart grid.Finally,through the simulation under the standard IEEE system s,the effectiveness of the false data attack detection model proposed in the paper is verified.