Generator-Level Transient Stability Assessment in Power System Based on Graph Deep Learning with Sparse Hybrid Pooling

Aimed at increasingly challenging operation conditions in modern power systems, online pre-fault transient stability assessment (TSA) acts as a significant tool to detect latent stability risks and provide abundant generator-level information for preventive controls. Distinguished from “system-level” to describe terms concerning the whole system, here “generator-level” describes those concerning a generator. Due to poor topology-related expressive power, existing deep learning-based TSA methods can hardly predict generator-level stability indexes, unless they adopt the generator dynamics during and after faults by time-domain simulation (TDS) as the model input. This makes it difficult to fully leverage the speed advantages of deep learning. In this paper, we propose a generator-level TSA (GTSA) scheme based on topology-oriented graph deep learning which no longer requires time-domain simulation to provide the dynamic features. It integrates two modules to extract the network-dominated interaction trends from only the steady-state information. A sparse Edge Contraction-based Attention Pooling (ECAP) scheme is designed to dynamically simplify the network structure by feature aggregation, where the generator-specific information and key area features are kept. A Global Attention Pooling (GAP) module works to generate the interaction features among generators across the system. Hence, the constructed ECAP&GAP-GTSA scheme can not only output the system stability category but also provide the dominant generators and inter-generator oscillation severity. The performance as well as interpretability and generalization of our scheme are validated on the IEEE 39-bus system and the IEEE 300-bus system under various operation topologies and generator scales. The averaging inference time of a sample on the IEEE 39-bus system and IEEE 300-bus system is merely 1/671 and 1/149 of that of TDS, while the accuracy reaches about 99%.