Enhancing trustworthiness in deep learning-based transient stability assessment: A multi-head attention model with dual-level interpretability

Applying deep learning (DL) methods to transient stability assessment (TSA) facilitates the rapid establishment of a direct connection between operational features and transient states. However, the inherent black-box nature of DL leads to decision-making mechanisms inexplicably, restricting their application in real-time TSA. In light of this, This paper proposes a TSA model based on multi-head attention DNN (MHA-DNN) with a dual-level interpretability framework. Firstly, a three-stage multi-head attention model integrated with a DNN is designed to capture spatial–temporal and feature correlations in transient processes. Subsequently, a dual-level transient stability assessment Explainer (DTSA-Explainer) is developed to provide insights into the overall decision-making process and the influence of individual TSA features through neuron conductance analysis and SHAP-based attribution. The influence patterns of the overall decision-making process and the input features on system stability can be revealed. Finally, a counterfactual preventive control is proposed to validate the effectiveness of the interpretability analysis, thereby improving the trustworthiness and transparency of the transient stability assessment results. Simulation results of the IEEE-39 bus system demonstrate that the proposed method offers superior TSA performance — outperforming benchmarks with 36.8% reductions in recall rates. Besides, the model maintains robust performance under topology variations and imbalanced datasets. Notably, the interpretation results from the DTSA-Explainer also enhance the understanding of system stability by aligning more closely with human intuition, offering a paradigm for interpretable DL in power systems. • Proposes a three-stage multi-head attention model for spatiotemporal correlation in transient processes. • Constructs a two-tier interpretability framework to explain decision mechanisms. • MHA-DNN outperforms baselines and enhances robustness to noise and imbalanced data. • Provides insights aligned with human intuition to assist in informed decision-making.