电力系统暂态稳定评估中神经网络的鲁棒性认证

Neural networks (NNs) can assess power system security rapidly and accurately, but they have limited robustness against small input perturbations that can lead to inaccurate predictions. Robustness certification can evaluate NNs' performance under perturbations, ensuring their credibility in practical applications. However, in transient stability assessment (TSA), the input data of NNs must comply with physical constraints rather than being subject to arbitrary perturbations. Additionally, even small input changes can affect transient stability. These two characteristics can cause inaccurate certification outcomes and make it challenging to directly apply traditional robustness certification methods in TSA. To address this, this paper introduces a robustness rate index considering physical feasibility and proposes a two-stage certification framework, where physical constraints of power systems are embedded to derive non-trivial robustness lower bounds while upper bounds are obtained by confirming the transient stability of optimized samples. Furthermore, the study explores model selection and adversarial training based on robustness certification. NNs are compared by robustness certification rate to select the most robust model among those with similar validation accuracy. Adversarial examples from certification are used to improve NNs' robustness. The proposed methods are validated on the New-England 10-machine system and a real-world regional system, demonstrating their effectiveness.