Vulnerability analysis, robustness verification, and mitigation strategy for machine learning-based power system stability assessment model under adversarial examples

Based on machine learning (ML) technique, the data-driven power system stability assessment has received significant research interests in recent years. Yet, the ML-based models may be vulnerable to the adversarial examples, which are very close to the original input but can lead to a different (wrong) assessment result. Taking short-term voltage stability (STVS) assessment problem as the case study, this paper firstly analyzes the vulnerability of the ML-based models under both the white-box and the black-box attack scenarios, where adversarial examples are generated to falsify the STVS assessment model into the wrong outputs without noticeable changes of the input values. Then, an empirical index is proposed to quantitatively measure the robustness of ML-based models under adversarial examples. After that, an adversarial training-based mitigation strategy is proposed to enhance the ML-based model against the adversarial examples under both the white-box and the black-box scenarios. Simulation results have clearly illustrated the threat of the adversarial examples to the ML-based models and verified the effectiveness of the proposed mitigation strategy.