Current deep learning architectures

particularly Transformer-based models

have demonstrated superior performance in power system transient stability assessment.However

conventional Transformers exhibit insufficient localized feature extraction capability for transient data characteristics.To address this limitation

we propose a Swin Transformer-based transient stability evaluation framework for power systems.By replacing the fixed block-based self-attention mechanism with a hybrid windowing approach that integrates non-overlapping local windows and shifted windows

our method enables synergistic computation of local-global attention patterns.This architecture effectively captures multi-grained features from power system transient data while improving predictive accuracy.Furthermore

our framework incorporates attention mechanism-guided key component identification through adaptive focus allocation

where attention weight distributions are analyzed to establish correlations with system instability modes.This not only enhances model interpretability but also optimizes computational efficiency.Notably

the proposed model exhibits linear computational complexity with respect to feature map dimensions

contrasting with the quadratic complexity inherent in traditional Transformers.Simulation results on the IEEE 10-machine 39-bus system demonstrate that our approach outperforms conventional deep learning and machine learning methods in stability assessment accuracy while maintaining lower computational costs compared to standard Transformer architectures.