融合时序卷积与多头自注意力的暂态电压稳定评估OA

To enhance the performance of transient voltage stability assessment models and improve the interpretability of assessment results,a method integrating temporal convolutional and a multi-head self-attention mechanism is pro-posed in this paper.First,to accurately assess the power system's transient voltage stability,the Transformer encoder is modified by embedding a temporal convolutional module,thus capturing the global and local information about elec-trical parameters throughout the transient process.Second,an adaptive threshold focal loss function is put forward to mitigate the adverse impact of sample imbalance on model training.Third,an interpretability analysis method based on the multi-head self-attention mechanism is employed to calculate attention weights across both the temporal and spatial dimensions,providing insights into the decision-making process of the assessment model.Finally,the proposed method is verified by simulations conducted on an IEEE-39 bus system and an IEEE-300 bus system,and results demonstrate that it achieves higher assessment accuracy,exhibits strong robustness,and offers interpretability.