A Hybrid Deep Learning Framework With Local-Global Feature Extraction for Intelligent Power System Stability Assessment

Transient simulation is the key to ensure the safe and stable operation of power systems. After large disturbances, power systems may experience transient rotor angle instability and short-term voltage instability, which share similar electrical characteristics but require different control strategies. Therefore, distinguishing the dominant instability mode (DIM) is crucial. In this paper, a novel hybrid deep learning (DL) framework is proposed to achieve accurate DIM identification by fully extracting local-global features from electricity data. In the proposed framework, an advanced graph neural network (GNN) optimized by random sampling and aggregation is constructed. It can better capture local features than traditional GNN and help to improve model generalization ability. Further, important global features are mined by employing transformer network with self-attention mechanism to improve DIM identification accuracy. In addition, the vital discrete fault features are also embedded into the neural networks to improve the performance. The proposed method improves upon existing DL models by providing a solution for integrating multi-level features extracted by different schemes and addressing the limitation of focusing on a single instability mode. Case studies conducted on an 8-machine 36- bus system and Northeast China Power Grid verify the superiority of the proposed method over other state-of-the-art DL methods.