Safety Assurance of AI-Enabled Power Cyber-Physical Systems: A Power-System-Centric Critical Review

Abstract

Artificial intelligence (AI) is increasingly used in power-system forecasting, stability assessment, control, protection support, and cyber-attack detection. When these models are embedded in power cyber-physical systems (Power CPS), however, prediction errors, distribution shifts, adversarial measurements, communication delays, and insufficient operator oversight may propagate into physical consequences such as reserve shortage, voltage or frequency violations, delayed protection response, or degraded restoration. This paper reviews safety assurance for AI-enabled Power CPS from a power-system operational perspective. Instead of treating AI safety as a generic property of machine-learning models, the review focuses on how AI decisions interact with SCADA/PMU measurements, automatic generation control, inverter-dominated operation, protection and emergency control, digital-twin validation, federated learning, and human supervision. The literature is organized around four questions: how unsafe AI behavior arises in grid operation; which mechanisms can improve robustness, uncertainty awareness, verification, and runtime monitoring; how model-level errors should be translated into power-system consequences; and what evidence is needed before AI-enabled functions can be deployed in safety-critical control-room environments. The review shows that current studies still rely heavily on offline accuracy, limited simulation scenarios, and loosely defined trustworthiness concepts, while practical deployment requires scenario-based stress testing, closed-loop validation, fallback logic, and auditable human-in-the-loop procedures. Open challenges are identified in power-system-specific benchmarks, scalable verification, runtime assurance for distributed control, and certification-oriented deployment.