Transient Stability Self-Adaptive Assessment Of Power System Based On Transfer Learning

With the proposal of the dual carbon goal,the pace of building a new power system is accelerating.Driven by the energy transition,the dynamic characteristics of the actual power system are becoming more and more complex,and the traditional transient stability assessment methods cannot meet the requirements of practical applications.The application of data-driven artificial intelligence technology to the transient stability of power system has become a development trend that combines the demand from the inside out and the drive from the outside to the inside.Relying on the National Key R&D Program(2018YFB0904500),this paper takes Central China Power Grid as an example to deeply study transient stability self-adaptive assessment methods of the power system based on transfer learning.The main research contents are as follows:(1)Due to the time-varying and repetitive characteristics of the grid structure,the training data of a single topology cannot cover all operating conditions,and the model needs to be updated within a limited time.Therefore,in order to efficiently generate samples required for real-time assessment under the premise of ensuring sample quality,a scene clustering method based on improved k-medoids algorithm facing Central China Power Grid is established.44 typical topological structures covering one year of Central China Power Grid are available.The status of generators and lines which shows their input or withdrawal in the system is used as the characteristics of the topology.And the similarity between systems is measured by cosine similarity.A dissimilarity matrix is introduced to improve the k-medoids algorithm,and the annual topology is divided into four categories.The validity of the clustering results is proved by the analysis of examples.(2)In order to meet the requirements of the actual system for the rapidity and accuracy of transient stability assessment,an improved DBN-based transient stability assessment model is constructed.First of all,in view of the large scale of the actual power grid and the large time cost of the simulation samples,a method of generating actual power grid samples based on the electrical betweenness screening of key lines is proposed.Time-domain simulation of all topologies is performed offline to generate samples for transient stability assessment,and trajectory cluster features are extracted input features of the model.Then,optimize the DBN network structure from three aspects: activation function,loss function and reverse fine-tuning method.The attention mechanism is fused to improve the information extraction ability of input features,and the optimal structure parameters of the network are automatically searched based on particle swarm optimization.The optimized DBN network is used to mine the mapping relationship between sample features and transient stability,and the DBN evaluation models for the same topology are trained respectively.Finally,the excellent performance of the evaluation model is verified from multiple perspectives through numerical examples,which also lays the foundation for subsequent transfer learning and sample enhancement.(3)Aiming at the problem that the performance of deep learning model decreases after the great change of power grid topology,combining the characteristics of actual power grid,the model migration and sample migration are combined to realize the adaptive evaluation of power system transient stability.And based on the above scene clustering results,two basic principles of sample transfer are proposed: scene matching principle and sample similarity principle.First,after the topology changes,the topology features of the new scene are extracted and matched with the historical scenes,and the class of the scene with the highest matching degree is used as the source domain for transfer.The structure of the source domain model and all parameters are fixed and transferred to the target domain.Then the samples in the new scene are generated online,and the samples in the source domain are transferred to the target domain according to the transfer principle and the specified transfer threshold,and together with the samples generated online,they constitute new training samples required for fine-tuning of the model.Finally,the new training samples are used to fine-tune the model as a whole,and a model suitable for transient stability assessment of the new scene is trained.The example results show that the proposed method can improve the prediction accuracy of the model when dealing with systems with large changes in topology,and can efficiently complete the process of sample generation and acquisition,taking the rapidity and accuracy of assessment into account.