Power System Transient Stability Assessment Based On Convolutional Neural Networks

Power system transient stability assessment(TSA)is one of the important issues that must be considered to ensure the safe and stable operation of the power system and improve the economic benefits of system operation.With the advent of the era of big data in electric power and the widespread application of wide area measurement system(WAMS),it has provided a rich source of data for TSA based on real-time response information of the system.The development of artificial intelligence and deep learning provides a new technical route for better mining the hidden mapping relationship between system response information and system steady state.How to make artificial intelligence and deep learning better serve the field of power systems has attracted continuous research attention among experts and scholars.Convolutional Neural Network(CNN)is a neural network model with deep architecture for deep learning.It has powerful feature expression capabilities and advantages in processing high-dimensional non-linear data.And it has been successfully applied in image recognition and other fields.However,the application in TSA is still in its infancy.Therefore,this dissertation has conducted exploration and in-depth research in this field of study.Based on the practical application requirements of transient stability analysis,suitable online real-time TSA methods focusing on the key problems that need to be solved in practical engineering applications are proposed.The main research contents of this dissertation are as follows:(1)Aiming at the problem of power system TSA,a TSA method based on a single CNN model is proposed.In terms of input features,27 geometric features based on the dynamic response trajectory cluster of generator power angles after fault removal are constructed as the wide-area fault features of the TSA model.In terms of evaluation model,the CNN model with strong processing capabilities for input features is selected.Based on the analysis of the basic principles of CNN,the effects of the size of convolution kernels,mini-batch,training iterations,and improved loss functions on the evaluation performance of a single CNN model are researched.The applicability and effectiveness of the proposed model are verified.Furthermore,considering the possibility of noise and incomplete WAMS information in actual online applications,the robustness of the proposed CNN TSA model is tested and analyzed.It is verified that the CNN TSA model based on the proposed trajectory cluster features has strongrobustness and can well meet the requirements of TSA of power systems.(2)Aiming at the problem that the existing models do not pay enough attention to the situation that the unstable samples in the prediction result are misdetecteded as stable samples,an integrated CNN model for TSA that takes into account the misdetection and false alarm is proposed.Multiple input feature sets are extracted based on the original measurements and the calculation of trajectory cluster features.Under each input feature set,several sets of CNN parameters with high assessment accuracy are selected,and an integrated CNN model is established through an ensemble learning strategy with probability average.Combining the optimization of the coefficient ratio of loss function weight,the binary classification threshold,and the ensemble learning strategy,a comprehensive evaluation criterion that takes into account misdetection and false alarm is proposed.So that the evaluation performance of the integrated CNN model can reduce the rate of misdetection of the unstable samples as much as possible while maintaining an acceptable false-alarm rate of stable samples,enhancing the practical application value of the model.(3)In order to take into account the speed and accuracy of the TSA model,a method based on the integrated CNN model is proposed.In the transient stability prediction,a two-stage transient stability prediction method combining a hierarchical real-time prediction method with credibility threshold optimization and an emergency control startup strategy based on multi-criteria fusion is proposed.This method can significantly reduce the false-alarm samples of actual startup emergency control and reduce or even eliminate misdetection of the unstable samples at the lowest cost,and improve the accuracy of transient stability prediction based on AI method.In the assessment of transient stability degree,by constructing regression prediction models for stability and instability degree,the transient stability and instability degrees are evaluated for the credible stable samples of the hierarchical prediction output and the credible unstable samples that meet the emergency control startup conditions,respectively.It further improves the precision of assessment and provides reference for the follow-up preventive control and emergency control measures.(4)Aiming at the problem of the lack of adaptability of the existing evaluation models,an adaptive evaluation method for transient stability of power systems based on transfer learning and CNN is proposed.This method consists of an effective scheme of CNN knowledge transfer learning and a method for generating minimum equilibrium sample set for transfer learning.It can greatly reduce the generation time of newsamples and the training time of model updates,and improve the adaptive speed of the evaluation model during online operation.First,a large number of transient stability samples generated offline are used to train a CNN-based pre-training model.The structure of network and the parameters of convolutional layers,pooling layers,and the fully connected layer of the pre-trained CNN model are kept unchanged,when the system operation mode and topology change greatly.A minimum equilibrium sample set is generated by combining variable step size and dichotomy method to retrain the classification layer parameters.Thus,the adaptive evaluation of the transient stability of the power system under different operation modes and topologies is realized.This method is not only applicable to the single CNN model,but also to the integrated CNN model.And it can be used in the hierarchical self-adaptive method proposed in this paper based on sliding time window input features to achieve online continuous self-adaptive TSA.