Protection-grade artificial intelligence for modern power systems

Modern power systems are increasingly challenged by inverter-based resources, fast converter controls, topology variability, hybrid AC/DC interfaces, digital substations, and evolving protection operating conditions. These developments weaken some assumptions underlying conventional protection settings and coordination, while increasing the need for adaptive supervision, uncertainty-aware decision support, and traceable engineering workflows. This paper proposes a protection-grade framework for integrating AI into modern power system protection as an augmentation layer rather than a substitute for deterministic relay logic. The framework first establishes a protection application taxonomy across assets, schemes, network levels, and operating timescales, and classifies AI functions according to decision authority, including primary-assistive, backup-assistive, supervisory, blocking/alarming, post-event analytical, and engineering-support roles. It then develops a protection-realistic data and infrastructure foundation based on disturbance records, PMU and micro-PMU streams, IEC 61850-based digital-substation measurements, asset-condition channels, synchronization requirements, labeling limitations, and simulation-to-field mismatch. Building on this foundation, the paper synthesizes protection-grade AI design patterns for regime recognition, supervisory recommendation, drift discovery, multimodal fusion, and hybrid or physics-informed modeling under practical constraints such as bounded latency, deterministic fallback, explainability, and traceability. The review further examines AI applications in adaptive relaying, element supervision, microgrids, wide-area protection intelligence, system integrity protection, engineering workflows, post-event analytics, and condition-informed protection enablement. A deployment-assurance and evidence-maturity assessment is introduced to compare research studies, HIL/RTDS validations, pilot projects, and field-oriented deployments beyond accuracy alone. The synthesis shows that AI evidence is currently stronger for supervisory, analytical, engineering-support, and condition-informed functions than for autonomous primary tripping. Accordingly, the paper identifies remaining gaps in benchmark datasets, degraded-data testing, uncertainty calibration, closed-loop validation, lifecycle governance, and utility-scale deployment, and outlines realistic future steps toward defensible protection-grade AI adoption.