Enhancing Smart Grid Cyber Resilience Against FDI Attacks Using Multi-Agent Recurrent DDPG

Digital substations (DSs) play a critical role in modern Energy and Power Electrical Systems (EPESs), enabling intelligent control, monitoring, and automation. With increased reliance on communication and sensing technologies, DSs are vulnerable to cyberattacks such as False Data Injection (FDI). An adversary may falsify transformer temperature readings, misleading protection mechanisms and resulting in incorrect disconnection actions. These false disconnections may disrupt power delivery, cause economic losses, and reduce equipment lifespan. To address these challenges, we propose a reinforcement learning-based approach for cyber protection of smart grids against false temperature data injection attacks. Specifically, this work designs a Long Short-Term Memory Deep Deterministic Policy Gradient (LSTM-DDPG) deep reinforcement learning algorithm that learns to detect normal patterns and responds to suspicious thermal patterns by dynamically adjusting disconnection decisions. The agents process sequential state features to differentiate between legitimate overload conditions and sudden anomalies caused by FDI attacks. We implement the proposed approach on the IEEE 30-bus distribution network using the Pandapower simulator. The experimental results indicate that the LSTM-DDPG controller outperforms conventional DDPG and DQN baselines, achieving a recall of 0.897, F1 of 0.945, precision of 1.00 and accuracy of 0.981 with a confidence interval of 95%. In addition, grid stability reaches up to 0.9815, 1.0, 1.0, 0.9926 with respect to the voltage stability score, transformer stability value, disconnection stability, and stability index, respectively. The proposed method led to fewer false disconnections, providing improved robustness against sensor manipulations.