Research On The Detection Of False Data In Power System Based On Machine Learning

With the development of the Internet of Things(Io T)technology,the grid intelligence will be improved more drastically,and in the process,the information attack will be more harmful than normal expectation.Currently,the boundary between physical and informational networks is increasingly blurred,and the power system is gradually evolving into a typical information-physical convergence system(CPS).Traditional state estimation assessment has been widely used in power systems,but there are vulnerabilities.False data injection attack(FDIA),which can construct false measurement data injected into the SCADA system without increasing the estimation residuals,successfully bypasses the detection module and leads to false state estimation results due to the distortion of the measurement data,which makes the stability of the system seriously affected,and even causes the local grid to be paralyzed in serious cases.Therefore,effective detection,identification and defense of false injection attacks have become a hot topic in current research on grid security.The methods for false data attack detection are studied in depth based on machine learning,and the following aspects are mainly studied:Firstly,based on the analysis of the theory of grid state estimation,the basic principle and construction method of the false data injection attack are studied.Meanwhile,the Matpower simulation toolkit is used to generate the normal abnormal data sample sets required for detection in two different standard test systems,IEEE14-bus and IEEE118-bus,to provide a data basis for building and training subsequent detection models.Secondly,to improve the false data detection accuracy,an improved algorithm of isolated forest algorithm combined with binary difference evolution without parameter variation is proposed for identifying false data.Since some tree structures with poor detection capability can reduce the detection accuracy,the binary difference evolution without parameter variation is used to find out the solution set with high accuracy and variance,and this combination is used to construct a high accuracy forest detection model.The effectiveness of the improved method is verified by simulating two standard nodes,IEEE-14 and IEEE-118.The results demonstrate that the improved detection method is significantly better than the traditional isolated forest detection method in terms of correctness,reaching 96% on the IEEE-14 node and 98% on the IEEE-118 node.Finally,a deep belief network-based false data detection method is proposed for grid topology changes.Since the circuit topology changes and the dimensionality of data increases,and the isolated forest is not effective in detecting high-dimensional data with low correct detection rate,a deep belief network that can extract useful information in high-dimensional data is chosen.The research results of this thesis are verified by simulating two standard nodes,IEEE-14 and IEEE-118.The results demonstrate that this method achieves a correct rate of 89% on the IEEE-14 node and 87% on the IEEE-118 node.