A transient stability assessment method for power systems incorporating residual networks and BiGRU-attention

Abstract

The traditional transient stability assessment (TSA) model for power systems has three disadvantages: capturing critical information during faults is difficult, aperiodic and oscillatory unstable conditions are not distinguished, and poor generalizability is exhibited by systems with high renewable energy penetration. To address these issues, a novel ResGRU architecture for TSA is proposed in this study. First, a residual neural network (ResNet) is used for deep feature extraction of transient information. Second, a bidirectional gated recurrent unit combined with a multi-attention mechanism (BiGRU-Attention) is used to establish temporal feature dependencies. Their combination constitutes a TSA framework based on the ResGRU architecture. This method predicts three transient conditions: oscillatory instability, aperiodic instability, and stability. The model was trained offline using stochastic gradient descent with a thermal restart (SGDR) optimization algorithm in the offline training phase. This significantly improves the generalizability of the model. Finally, simulation tests on IEEE 145-bus and 39-bus systems confirmed that the proposed method has higher adaptability, accuracy, scalability, and rapidity than the conventional TSA approach. The proposed model also has superior robustness for PMU incomplete configurations, PMU noisy data, and packet loss.