Scholarlink

<p>Dynamic state estimation (DSE) is important for real-time monitoring and secure operation of power systems, but it is affected by nonlinear dynamics, asynchronous PMU/SCADA measurements, non-Gaussian noise, and false data injection attacks (FDIAs). This paper proposes a twin-delayed deep deterministic policy gradient (TD3)-assisted cubature Kalman filter (CKF) for robust multi-rate DSE. The TD3 agent adaptively scales the effective measurement covariance using residual information, innovation statistics, measurement arrival status, and previous actions, thereby regulating the Kalman gain online without changing the CKF structure. Simulations on the IEEE 39-bus system compare fixed-covariance CKF, innovation-adaptive CKF (IAE-CKF), and TD3-CKF under clean, non-Gaussian, sparse additive FDIA, and simplified stealthy FDIA cases. An additional IEEE 118-bus test is also conducted to evaluate scalability. The results show that TD3-CKF does not uniformly outperform all baselines, but it achieves competitive accuracy and improves robustness in several corrupted-measurement scenarios. Sensitivity and scalability tests further indicate that the learned covariance policy has useful adaptability, although its advantage is scenario-dependent and model-consistent stealthy attacks remain challenging.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep reinforcement learning–assisted cubature Kalman filtering for robust multi-rate dynamic state estimation under false data injection attacks

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Mingxing Qiao,
  • Dongjian Huang,
  • Wenyuan Wang,
  • Wenxu Yan

摘要

Dynamic state estimation (DSE) is important for real-time monitoring and secure operation of power systems, but it is affected by nonlinear dynamics, asynchronous PMU/SCADA measurements, non-Gaussian noise, and false data injection attacks (FDIAs). This paper proposes a twin-delayed deep deterministic policy gradient (TD3)-assisted cubature Kalman filter (CKF) for robust multi-rate DSE. The TD3 agent adaptively scales the effective measurement covariance using residual information, innovation statistics, measurement arrival status, and previous actions, thereby regulating the Kalman gain online without changing the CKF structure. Simulations on the IEEE 39-bus system compare fixed-covariance CKF, innovation-adaptive CKF (IAE-CKF), and TD3-CKF under clean, non-Gaussian, sparse additive FDIA, and simplified stealthy FDIA cases. An additional IEEE 118-bus test is also conducted to evaluate scalability. The results show that TD3-CKF does not uniformly outperform all baselines, but it achieves competitive accuracy and improves robustness in several corrupted-measurement scenarios. Sensitivity and scalability tests further indicate that the learned covariance policy has useful adaptability, although its advantage is scenario-dependent and model-consistent stealthy attacks remain challenging.