Research On Security Monitoring With Deep Learning In Smart Grids

With the increasing number of attacks in smart grids,the security performance decreases.To improve the security performance,we propose an attack behavior detection architecture.It can distinguish attack behaviors into known and unknown,then,based on Deep Learning(DL),it extracts representative characteristics of different attack types,thus ensure the accuracy of attack behavior recognition.The main contributions of this article are as follows:(1)To detect abnormal behaviors caused by a large number of attack behaviors and long data fragments in the smart grid,we propose a behavior detection method based on GAN.The proposed method is mainly based on GAN to optimize the learning strategy of generator and discriminator.Among them,in the discriminator,the loss function is converged by maximizing the root mean square error of the attack,and the linear activation function is used as a regulating function.Therefore,the improved GAN can take advantage of the characteristics of data enhancement in the original model,and can ensure the stability of attack behavior recognition,then effectively detects the attack information in the smart grid.(2)In order to improve the success rate of the attack,the detected abnormal behavior increases its concealment.We proposed Deep Belief Netwoks(DBN)to automatically generate and select features,reducing the analysis cycle and ensureing accuracy.Specifically,based on the improved DBN,a non-linear iterative algorithm under constrained features is proposed.The proposed algorithm changes the iterative input data based on the stability of the output result,and converts known attack conditions into constraints,so as to solve the attack vector.Then,we use DBN to perform deep analysis on each attack vector element of the output.The proposed method can effectively avoid noise interference,thereby improving the recognition accuracy of hidden behaviors.(3)For unknown behavior detection,we translate it into representative behavior characteristics from unknown attack behaviors.In order to solve this problem,this thesis proposes a method based on support vector machine(SVM)for unknown attack behavior feature extraction.Using the weight vector of each parameter in each layer of the neural network(NN).After the regularization is solved,the non-initialization processing in the behavior vector is automatically deleted,and the features are classified and sorted,then multiple feature sets are generated.The proposed method can effectively extract representative behavior characteristics in unknown attacks and improve the detection accuracy of unknown behaviors in smart grids.