A transfer learning-based graph convolutional network for dynamic security assessment considering loss of synchronism of wind turbines and unknown faults

Abstract

Pre-fault dynamic security assessment (DSA) is essential for the safe operation of power systems. Pre-fault DSA methods that utilize deep learning techniques have been successfully implemented and have shown promising results. However, these methods face challenges in real power systems, such as unknown faults and the increasing integration of power electronics-based units. Adding these units changes the system dynamics and introduces new stability problems, such as the loss of synchronism that current methods cannot analyze. In practical applications, new faults may arise that are not present in the training database, which can decrease the accuracy of the online DSA model. To tackle these challenges, this paper introduces a new dynamic security index that considers the effects of loss of synchronism in power electronics-based units on DSA. Also, a graph convolutional network (GCN)-based model is developed to improve DSA accuracy by incorporating the topological information of the power system in the form of an adjacency matrix. To address the issue of unknown faults, this paper uses transfer learning based on full fine-tuning to adapt a pre-trained GCN model to a different but related unknown fault. This approach eliminates the need for a large number of labeled examples for new faults and ensures efficient transfer of the model to new faults with a small database. Case studies are conducted on a modified IEEE 39-bus system to investigate the impact of power electronics-based units' penetration on dynamic security and the model's ability to transfer knowledge for unknown faults. The results from various evaluation indicators demonstrate the effectiveness of the proposed model.