Physical information-guided training method for transient voltage stability assessment

Deep learning models provide a new solution to power system transient voltage stability assessment (TVSA) problems but encounter tricky challenges in result interpretability and process controllability. In this paper, a guided training method incorporating physical information of the critical nodes is proposed for interpretable and controllable TVSA. First, a Transformer-based TVSA model is proposed to interpret the TVSA rules through learned attention distributions. Second, a voltage sag severity index is utilized to mark the critical nodes that imply the information of different voltage instability patterns. Third, an attention guidance term is proposed and integrated into the Transformer model to accurately capture critical features of targeted voltage instability patterns. Then, the TVSA rules can be further optimized and the misprediction of important voltage instability patterns can be mitigated, thereby improving the controllability of the TVSA model. Additionally, a logic factor is dedicatedly designed to reduce the cost of model-guided training. The effectiveness of the proposed method is verified in the IEEE 39-bus system and the realistic regional power system.