Deep-Learning-State-Estimation-Aided Detection Framework against False Data Injection Attacks
Bo Liu, Yichen Henry Liu, Xuebo Liu, Hongyu Wu · 2024
False data injection (FDI) attacks can bypass bad data detection and mislead state estimation (SE), resulting in economic losses and security issues. Existing FDI attacks consider the spatial correlation without considering the temporal correlations. Therefore, FDI attacks can correctly mislead the traditional Weighted Least Square SE (WLS-SE) with desired voltage incremental, but hard to accurately mislead the deep-learning-based SE to the desired malicious voltage. This paper first proposes a long-short-term-memory-based state estimator (LSTM-SE), and then proposes a novel deep-learning-SE-aided (DLSEA) attack detection framework. The proposed detection framework utilizes the voltage estimation difference (VED) between the WLS-SE and LSTM-SE to detect the attacks. A fully connected neural network is utilized to classify the VED values for determining either normal system conditions or under cyberattacks. Numerical results in the IEEE 14-bus and IEEE 118-bus systems show the proposed LSTM-SE can approximately estimate the true voltage under FDI attacks, and the proposed detection framework can detect FDI attacks with 0.99 accuracy. We further evaluate the impact of noises on the performance of LSTM-SE and DLSEA.