A Learning-Based Solution for an Adversarial Repeated Game in Cyber-Physical Power Systems

Area of Science:

• Cyber-physical systems
• Power system security
• Game theory applications

Background:

• Increasing complexity of cyber-physical power systems elevates malfunction risks.
• Existing smart grid security models struggle to replicate real-world adversarial events accurately.

Purpose of the Study:

• To formulate a repeated game model for simulating realistic interactions between adversaries in modern electric power systems.
• To analyze optimal action strategies for attackers and defenders under various environmental conditions.
• To develop a reinforcement learning-based solution for achieving favorable outcomes in smart grid security games.

Main Methods:

• Formulation of a repeated game to mimic real-life adversarial interactions in power systems.
• Application of a reinforcement learning algorithm to determine optimal strategies for players.
• Inclusion of factors like costs, budgets, and player strengths to enhance game realism.
• Evaluation using the IEEE 39 bus system and simulation in PowerWorld.

Main Results:

• The repeated game model successfully replicates real-world adversarial dynamics in power systems.
• Reinforcement learning enables active learning of attack and defense strategies for power transmission lines.
• Analysis shows favorable outcomes for players based on mixed strategy payoffs, considering various influencing factors.
• Simulations demonstrate the impact of learned strategies on power system stability and resilience.

Conclusions:

• The proposed game-theoretic approach with reinforcement learning offers a more realistic simulation of smart grid security.
• Active learning of strategies by both attackers and defenders enhances the robustness of power systems against cyber threats.
• The model provides valuable insights into optimizing defense mechanisms and mitigating power loss in complex power grids.