An imbalanced deep learning framework for pre-fault flexible multi-zone dynamic security assessment via transfer learning based graph convolutional network

Abstract

• A data augmentation model based on CTGAN is implemented in the minority class to balance the database in the pre-fault DSA. Unlike traditional oversampling methods that generate samples by simple linear sampling, this technique creates a balanced and realistic database from the original database. • A GCN-based DSA model with power system zoning is proposed. By capturing the system's topological structure, GCN performs spatial dependency modeling, which enhances the model's accuracy. Also, the power system zoning and training a DSA model for each zone avoid unnecessary updates for some zones when the topology is changed, which reduces computational costs. • A flexible update approach is proposed, using the MMD index as a trigger to activate the update. According to the MMD value, different transmission approaches are selected adaptively for each topology and each region of the power system. This increases effectiveness and reduces the time needed to update. The proposed transfer approaches use a pre-trained model that allows updating with a small database. Deep learning (DL)-based pre-fault dynamic security assessment (DSA) methods have shown promising results. However, DL-based DSA faces challenges related to model robustness against topology changes and database imbalances. Although an accurate model can be trained for a particular topology, it often does not work for other topologies and requires updating. Also, most existing methods consider a balanced labeled database to train a DL-based model optimally. Meanwhile, as modern power systems become more stable, the number of secure operating conditions is far more than that of insecure ones. This causes the training database to be imbalanced and reduces the model's efficiency. This paper addresses these two challenges with the help of transfer learning (TL) and conditional tabular generative adversarial networks (CTGAN). This paper first generates synthetic data using CTGAN, which helps create a balanced and representative training database to combat the negative effects of an imbalanced database. Then, the power system is zoned, and for each zone, a model based on a graph convolutional network (GCN) is trained with a balanced database to achieve better performance. The GCN-based model improves evaluation accuracy using power system topological information as an adjacency matrix. Since the impact of topology changes on fault behavior is not the same throughout the power system, power system zoning prevents unnecessary updates to some zones. Finally, when the power system topology changes, using maximum mean difference (MMD), the zones that need to be updated are determined, and the transfer approach for these zones is selected. In this paper, for high MMD, the fine-tuning approach of the entire model is employed, and for low MMD, the fine-tuning approach of the last two dense layers is employed. Using two different transfer approaches based on the MMD value reduces model update time and ensures effective performance. The proposed model was implemented on the IEEE 118-bus system and analyzed based on various indicators, and the results show its effectiveness.

Topics & Concepts