Research On The Evaluation Method Of Power Grid State And System Stability Based On Deep Learning And RMT

Now the interconnection scale of the power grid and its operating characteristics has been gradually complicated.Due to the random fluctuation and complex coupling of the power grid,it has become increasingly difficult to achieve the operating state analysis and transient stability assessment in power systems quickly and accurately.Recently,the wide-area measurement system(WAMS),information communication technology(ICT)for 5G(5th-Generation)as well as data-driven technologies such as big data and artificial intelligence(AI)have been rapidly evolving.Although the mass data being generated is growing exponentially,data-driven technologies also provide more advanced technology for data processing and value application.On the one hand,random matrix theory(RMT)is a universal big data analysis methodology,which can understand the behavioral characteristics of the complex system from a high-dimensional perspective.On the other hand,deep learning(DL)is a pattern analysis approach closed to AI,which can learn the intrinsic rules and level of detail on sample data to implement complex learning tasks such as classification.Therefore,there is important theoretical significance and application potential to carry out the researches on the operating state analysis and transient stability assessment in power grids based on DL and RMT.Firstly,this paper introduces the research background of the subject,and the research status at home and abroad is reviewed.Furthermore,the basic principle of both RMT and DL is introduced.Then,the data preprocessing method suitable to RMT and the structure and train methods of DL model can be given,respectively.Finally,it mainly analyzes the applications for power grid state analysis and transient stability assessment(TSA).Second,the maximum eigenvalue of the sample covariance matrix(MESCM)-based the method and principle in RMT,a novel threshold model with convolutional neural network(CNN)suitable for abnormal load dynamic evaluation of power grids is proposed.It can be achieved the detection of abnormal load in power grids with large load fluctuation using MESCM.Then,the results in comparison with those results from the traditional threshold model by an IEEE145-bus 50-machine system indicate that proposed model is more adaptable and accurate for the judgment of abnormal MESCM index.Thirdly,an RMT-based evaluation method for unbalanced disturbance events in the power grid using the Spiked covariance model and phase transition(PT)phenomenon is proposed.Based on the MESCM method,the proposed method could be implemented for the identification and location of unbalance disturbance events employing the Spiked eigenvalues,Spiked eigenvectors as well as their PT phenomenon.The case studies have been carried on an IEEE 118-bus and 177-line system and a real 872-bus and 1840-line distribution network system in Germany,which involve unbalanced disturbance events such as load variations and line faults.The comparisons between the results from the proposed method and the traditional RMT-based method indicate that it is valid and efficient.Finally,according to the unique multi-channel mechanism of multi-channel convolutional neural network(MCNN),this paper proposes an MCNN-based multiple feature fusion method(MFF-CNN)for the transient stability prediction(TSP)in power grids.Then,TSP on both stability margin(SM)and stability status(SS)could be achieved employing the trained MFF-CNN model.By an IEEE 50-machine and145-bus system,the comparisons among the results from the proposed model and the traditional shallow and deep learning-based methods demonstrate that it is not only valid,but also superior performances in TSP.In addition,the time-series TSP for identifying the stability boundary samples is further achieved by estimating the confidence interval of the prediction results of proposed model.