Vulnerability and Impact of Machine Learning-Based Inertia Forecasting Under Cost-Oriented Data Integrity Attack
Vulnerability and Impact of Machine Learning-Based Inertia Forecasting Under Cost-Oriented Data Integrity Attack
复制标题
DOI：
10.1109/tsg.2022.3207517
复制
发表时间：
2023-05
影响因子：
9.6
通讯作者：
Yan Chen;Mingyang Sun;Zhongda Chu;Simon Camal;G. Kariniotakis;F. Teng
中科院分区：
文献类型：
--
作者：
Yan Chen;Mingyang Sun;Zhongda Chu;Simon Camal;G. Kariniotakis;F. Teng
随着可再生能源的日益普及，电力系统正面临着前所未有的低惯性水平挑战。系统通过惯性响应抵御干扰和功率不平衡的固有能力下降，因此，系统操作员需要做出更快、更有效的调度操作。作为最有前途的解决方案之一，机器学习（ML）方法已经被研究并用于实现有效的惯性预测，并且具有相当的准确性。然而，面对日益增长的网络攻击威胁，它还没有认识到自己的脆弱性。为此，本文提出了一个方法框架来探索基于ml的惯性预测模型的脆弱性，并特别关注数据完整性攻击。特别是，首次提出了一种以成本为导向的虚假数据注入攻击，其主要目标是通过最小化摄动前和摄动后惯性预测之间的差异，在显著提高系统运行成本的同时保持攻击的隐秘性。此外，我们提出了基于ml的惯性预测模型的4个脆弱性评估指标。通过对GB电力系统的案例研究，证明了基于ml的惯性预测模型的脆弱性和影响，以及所提出的以成本为导向的数据完整性攻击的隐蔽性和可移植性。
With the increasing penetration of renewables, the power system is facing unprecedented challenges of low-inertia levels. The inherent ability of the system to defense disturbance and power imbalance through inertia response is degraded, and thus, system operators need to make faster and more efficient scheduling operations. As one of the most promising solutions, machine learning (ML) methods have been investigated and employed to realize effective inertia forecasting with considerable accuracy. Nevertheless, it is yet to understand its vulnerability with the growing threat of cyberattacks. To this end, this paper proposes a methodological framework to explore the vulnerability of ML-based inertia forecasting models, with a special focus on data integrity attacks. In particular, a cost-oriented false data injection attack is proposed, for the first time, with the primary objective to significantly increase the system operation cost while retaining the stealthiness of the attack via minimizing the differences between the pre-perturbed and after-perturbed inertia forecasts. Moreover, we propose four vulnerability assessment metrics for the ML-based inertia forecasting models. Case studies on the GB power system demonstrate the vulnerability and impact of the ML-based inertia forecasting models, as well as the stealthiness and transferability of the proposed cost-oriented data integrity attacks.