QSTAformer: A Quantum-Enhanced Transformer for Robust Short-Term Voltage Stability Assessment against Adversarial Attacks

Short-term voltage stability assessment (STVSA) is critical for secure power system operation. While classical machine learning-based methods have demonstrated strong performance, they still face challenges in robustness under adversarial conditions. This paper proposes QSTAformer-a tailored quantum-enhanced Transformer architecture that embeds parameterized quantum circuits (PQCs) into attention mechanisms-for robust and efficient STVSA. A dedicated adversarial training strategy is developed to defend against both white-box and gray-box attacks. Furthermore, diverse PQC architectures are benchmarked to explore trade-offs between expressiveness, convergence, and efficiency. To the best of our knowledge, this is the first work to systematically investigate the adversarial vulnerability of quantum machine learning-based STVSA. Case studies on the IEEE 39-bus system demonstrate that QSTAformer achieves competitive accuracy, reduced complexity, and stronger robustness, underscoring its potential for secure and scalable STVSA under adversarial conditions.

💡 Research Summary

This paper addresses the critical challenge of Short-Term Voltage Stability Assessment (STVSA) in modern power systems, which are increasingly complex due to high penetration of renewable energy and power electronic devices. While data-driven machine learning methods have shown promise, they often suffer from high computational costs, vulnerability to adversarial attacks, and difficulties in modeling highly nonlinear dynamics. To overcome these limitations, the authors propose “QSTAformer,” a novel hybrid quantum-classical neural network architecture designed for robust and efficient STVSA.

The core innovation of QSTAformer lies in its integration of Parameterized Quantum Circuits (PQCs) into the self-attention mechanism of a Transformer model. This design leverages the theoretical advantages of quantum computing, such as accelerated inner product calculations in high-dimensional spaces and enhanced representational power through superposition and entanglement, to model complex power system dynamics more efficiently than purely classical counterparts. The paper provides a detailed benchmarking study of diverse PQC architectures (e.g., basic entangling layers, strongly entangled circuits) to explore the trade-offs between expressiveness, training convergence speed, and computational efficiency, offering practical guidelines for circuit design.

A significant and novel contribution of this work is the first systematic investigation into the adversarial vulnerability of Quantum Machine Learning (QML) models in the context of power system security. Recognizing that QML models are not inherently immune to threats, the authors evaluate QSTAformer under both white-box (full model knowledge) and gray-box (partial knowledge) attack scenarios using advanced techniques like Momentum Iterative FGSM (MI-FGSM) and Projected Gradient Descent (PGD). To counter these threats, they develop a dedicated adversarial training strategy that incorporates adversarial examples during the model’s training phase, significantly enhancing its robustness against such manipulations.

The proposed framework is comprehensive, incorporating auxiliary modules for data preparation. It utilizes a Semi-Supervised Fuzzy C-Means (SFCM) algorithm for soft-labeling PMU time-series data, which helps in scenarios with limited labeled data. Furthermore, a Least Squares Generative Adversarial Network (LSGAN) is employed for data augmentation to balance the dataset and improve the model’s generalization capability. These components feed into the main QSTAformer model for final stability classification.

Extensive case studies are conducted on the IEEE 39-bus power system. The experimental results demonstrate that QSTAformer achieves competitive classification accuracy compared to state-of-the-art classical models (e.g., Bi-directional LSTM, Graph Attention Networks) and existing quantum-based methods under normal operating conditions. More importantly, it exhibits superior robustness, maintaining significantly higher accuracy when subjected to adversarial attacks. Additionally, the model shows reduced parametric complexity and computational overhead, highlighting its efficiency. The study concludes that QSTAformer presents a promising, secure, and scalable solution for real-world STVSA applications, effectively bridging the gap between advanced quantum-inspired algorithms and the stringent reliability requirements of cyber-physical power systems.