Optimal power flow (OPF) is a foundational analysis in power system operation, and inaccuracies in OPF decision-making can result in significant operational and security consequences. With the increasing adoption of machine-learning-based OPF surrogate models for real-time and large-scale applications, power system operation has become more susceptible to cyber threats, including data manipulation and false data injection attacks. Explainable artificial intelligence (XAI), particularly Shapley value-based approaches, has been employed to quantify the influence of input features on OPF outputs. Features exhibiting high influence are commonly interpreted as vulnerable. However, this interpretation neglects the role of attack detection mechanisms in the system, which typically monitor input-to-input relationships to detect anomalous behavior. Consequently, highly influential features may be readily detectable and may not represent the most vulnerable attack surfaces. To address this limitation, we propose a defense-aware explainable AI framework that explicitly incorporates attack detectability alongside operational impact. The framework integrates an OPF surrogate model representing input-to-output dependencies with a vanilla autoencoder representing input-to-input dependencies. Shapley-based explainability is applied to both models to construct a feature-level vulnerability metric. Empirical evaluation on IEEE test systems demonstrates that the proposed framework more accurately identifies security-relevant vulnerabilities than conventional OPF-surrogate-model-only explainability approaches.

翻译：暂无翻译