Research On Power System Fault Treatment Method Based On Deep Learning

Power system failure treatment,the primary measure to ensure the safety and stability of the power system operation,has always been a critical issue during the construction and development process of our country’s power grid.With the booming of the third-generation artificial intelligence wave,as typified by deep learning(DL),the study of smart fault treatment solutions can adapt to the high dimension,correlation and uncertainty in the current large complex power grid,which provide a new insight into the power system operation.To his end,this paper focuses on the research on power system failure treatment methods on the basis of the DL technique.The main research work and contributions are summarized as follows:A novel intelligent method for internal and external fault diagnosis of transmission line based on single-ended traveling wave signal is proposed.The similarities and differences of single-ended fault traveling wave signal features at different fault locations are analyzed on the basis of the traveling wave transmission principle.The deep classification neural network is constructed by Gated Recurrent Unit.The network input is set as short sequences,which is obtained by dividing the single-ended fault traveling wave signal according to the line length.The network output is the probability of three type fault locations(reverse external fault,internal fault and forward external fault).The neural network training process is adopted Adam algorithm.The results of case studies show that the proposed method can accurately distinguish the internal and external faults of the transmission line,which has highly robust at the beginning and end of the line and strong ability of anti-noise.A novel quick calculation scheme of backup protection online setting based on Generative Adversarial Networks(GAN)is presented.A conditional GAN based on Wasserstein distance is constructed.It uses system operation modes as conditional labels and transforms the coordination relationship of backup protection into matrix indices to form the backup protection pair matrix as real samples for GAN training.The principal part of neural network is built by convolutional neural network to realize parallel computing.The results of case studies show that the proposed scheme that utilizes GAN to learn the coordination relationship of backup protection setting values in real samples,can give the corresponding backup protection pair matrix according to different operating modes and thus realizes the protection backup online setting quick calculation.A novel method for optimal allocation of Superconducting Fault Current Limiters(SFCL)using improved Deep Reinforcement Learning is advanced.A model based on the Markov Decision Process is established to convert the SFCL allocation into a Reinforcement Learning problem.The improved Q-learning algorithm is adopted to solve the problem,which save computation resources and enhance the convergence performance.A sensitivity-based greedy policy is presented to improve the action selection strategy in Reinforcement Learning,which can dramatically reduce the early inefficient optimization.The results of case studies show that the proposed method has excellent optimization ability and efficient convergence performance,which can solve the SFCL allocation problem in the large-scale power system.