| Nowadays,the modern power system is developing towards the direction of intelligence.A large number of intelligent devices,such as intelligent meters and sensors,promote the power system in power generation,transformation,transmission and distribution way,making the smart grid a typical network physical system,which combines the physical power transmission system and computer network.In smart grid,the Supervisory Control and Data Acquisition system(SCADA)collects the data sent by outfield devices through the network in real time,analyzes it and reports the collected information to the control center,which adjusts the power generation and distribution of the smart grid according to the information.While enjoying the convenience brought by smart grid,smart grid is more vulnerable to attack by hackers due to the use of a large number of smart devices and data transmission through the network.The most obvious is to attack the smart grid by premeditated tampering with the data of intelligent devices,which is called false data injection attack(FDIA).In order to guarantee safe power service operation,it is very important for smart grid detection.This thesis studies the problem of false data injection attack detection in smart grid,and the specific contents are as follows:1)This thesis summarizes the existing detection methods of FDIA with analyzing and comparing,and the future research direction is discussed.At present,the detection methods of FDIA in this field are mainly divided into two methods based on system model and data drive.At the same time,detection methods based on system model can be divided into estimation-based detection methods and other detection methods based on system model.Detection methods based on data drive can be divided into detection methods based on machine learning and other detection methods based on data drive.Both algorithms have advantages and disadvantages.In addition,the future research direction is discussed based on the existing detection methods and theories in other fields.2)A detection scheme based on Kalman Filter and Recurrent Neural Network(KFRNN)is proposed.Considering the smart grid data naturally contains linear and nonlinear components,corresponding prediction models(Kalman Filter for linear model and Recurrent Neural Network for nonlinear model)is used for parallel analysis in this thesis,then the results of two base learners are automatically assigned weights to output the final prediction result by using ensemble learning.Secondly,the prediction result is judged by the threshold obtained from the cumulative distribution curve of the sum of square errors between the predicted value and the actual value in the clean data set.Compared with other schemes,it shows higher detection probability and better detection result in the false data injection attack.3)In the real environment,the attackers who carry out the FDIA usually do not attack in the peak form,and most of them carry out random attacks to ensure sparse attacks.To solve this problem,a detection scheme based on multiple models and data redundancy(MMDR)is proposed.Firstly,one-dimensional data received by the SCADA are mapped to a two-dimensional matrix according to the topology of smart grid to increase the redundancy of data,and the two-dimensional matrix also contains the spatial characteristics of the smart grid.Secondly,we try to eliminate the injected false data by randomly deleting a certain row or column of two-dimensional data.Because deleting some data will miss some smart grid information,this thesis considers to output multiple prediction results by constructing multiple models for the negative impact of information loss.Then,the output results are divided into two categories by classification mechanism.The average value of the category with a large amount of data is taken as the final prediction result,and the threshold value obtained by the cumulative distribution curve of the sum of square error is used to determine whether there is false data injection attack.It is compared with supervised model based scheme and non-multi-model scheme in IEEE 14 buses. |
