.png){kind=icon}
.png){kind=icon}
Return
Explainable deep learning method for power system stability evaluation with incomplete voltage data based on transfer learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.measurement.2025.116781.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Real-time voltage assessment is critical for power system fault diagnosis. However, traditional deep learning methods often suffer from deficiencies in prediction accuracy and interpretability when handling missing voltage data, which limits their practical applicability. To address these challenges, this paper proposes the Interpretable Wavelet-Visual Transformer Model (X-WaveVT). By integrating wavelet transform, vision transformer (ViT), and transfer learning, X-WaveVT provides an effective solution for incomplete voltage data processing and power system stability assessment. The proposed method extracts multi-scale features from voltage data using wavelet transform and leverages the visual transformer to efficiently capture scale-space correlations, accurately distinguishing interference and noise caused by missing data. This enhances the model's ability to learn and utilize multi-scale features. When faced with severe data loss and limited samples, transfer learning fine-tunes a pre- trained model, effectively addressing data scarcity and significantly improving prediction accuracy. Furthermore, by incorporating a novel voltage time series driven Class Activation Mapping (CAM) technique, the model's decision process is visualized, enhancing interpretability and credibility. This visualization supports the extraction of key features and provides insights into the model's decision-making process. The experimental results show that the proposed method delivers outstanding performance in voltage stability assessment under full data conditions. Furthermore, as challenges such as data loss, sample insufficiency, and class imbalance become more severe, X-WaveVT exhibits a growing performance advantage over existing models, underscoring its stability and robustness in complex power systems. Additionally, the model's visualization capability not only enhances trust in its predictions but also offers valuable insights for real-time voltage assessment and power system optimization, underscoring its significant application potential and practical value.
Keywords:
Real-time voltage assessment
Missing voltage data
Explainable wavelet vision transformer
Transfer learning
Class activation score