Generative AI Enabled Robust Sensor Placement in Cyber-Physical Power Systems: A Graph Diffusion Approach

TL;DR

The paper tackles robust sensor placement in interdependent CPPS by formulating a joint physical-cyber optimization problem and proving its NP-hardness. It introduces the Experience Feedback Graph Diffusion (EFGD) algorithm, a diffusion-based graph-generation method augmented with cross-entropy gradient and trajectory-level RLHF feedback to accelerate convergence and improve final rewards. The framework leverages the LNSPL model for reliable communications, the Fiedler value and Cheeger-inequality bounds to quantify cyber-layer robustness, and three physical-layer anomaly detectors to quantify detection performance. Empirical results on the IEEE 118-bus system show EFGD achieving faster convergence (about 18.9% faster) and higher average rewards (up to ~22.9% over DDPO and ~19.6% over GDPO) than strong baselines, while generating sensor placements that preserve network connectivity under link failures. Overall, the work advances secure, reliable CPPS operation by jointly optimizing sensor deployment and cyber-layer resilience using a diffusion-based learning paradigm.

Abstract

With advancements in physical power systems and network technologies, integrated Cyber-Physical Power Systems (CPPS) have significantly enhanced system monitoring and control efficiency and reliability. This integration, however, introduces complex challenges in designing coherent CPPS, particularly as few studies concurrently address the deployment of physical layers and communication connections in the cyber layer. This paper addresses these challenges by proposing a framework for robust sensor placement to optimize anomaly detection in the physical layer and enhance communication resilience in the cyber layer. We model the CPPS as an interdependent network via a graph, allowing for simultaneous consideration of both layers. Then, we adopt the Log-normal Shadowing Path Loss (LNSPL) model to ensure reliable data transmission. Additionally, we leverage the Fiedler value to measure graph resilience against line failures and three anomaly detectors to fortify system safety. However, the optimization problem is NP-hard. Therefore, we introduce the Experience Feedback Graph Diffusion (EFGD) algorithm, which utilizes a diffusion process to generate optimal sensor placement strategies. This algorithm incorporates cross-entropy gradient and experience feedback mechanisms to expedite convergence and generate higher reward strategies. Extensive simulations demonstrate that the EFGD algorithm enhances model convergence by 18.9% over existing graph diffusion methods and improves average reward by 22.90% compared to Denoising Diffusion Policy Optimization (DDPO) and 19.57% compared to Graph Diffusion Policy Optimization (GDPO), thereby significantly bolstering the robustness and reliability of CPPS operations.

Figures (9)

• Figure 1: The system model of interdependent CPPS. Part A. The illustration of the physical layer grid with edge anomaly. Part B. The cyber layer network with link failures. Part C. The control center processes data and controls the CPPS.
• Figure 2: The illustration of the optimization objections. Part A. The proposed CPPS system based on IEEE 9 bus system. Part B. The anomaly detectors focusing on the safety of the system. Part C. The communication quality based on LNSPL model targeting the reliability. Part D. The Fiedler value of different graphs addressing the robustness of the system.
• Figure 3: The illustration of EFGD framework. Part A. The denoising network structure based on graph transformer architecture. Part B. The graph generation process based on the denoising graph diffusion model. Part C. The proposed experience feedback module.
• Figure 4: The average reward (AvgReward) of EFGD and other baselines over training epochs.
• Figure 5: The average reward (AvgReward) of EFGD and other baselines in the inference stage.

Theorems & Definitions (2)

• Proposition 1
• proof