TRUSTWORTHY ARTIFICIAL INTELLIGENCE FOR INVERTER-DOMINATED POWER SYSTEMS: A SYSTEMATIC REVIEW OF ROBUSTNESS, GENERALIZATION, AND REAL-TIME DEPLOYMENT

Abstract

The integration of inverter-based resources into modern power systems introduces fundamental challenges to conventional protection schemes, as low inertia, limited fault current, and operational variability render traditional methods inadequate. Artificial intelligence offers adaptive solutions for fault detection, cyberattack mitigation, and stability assessment, yet concerns about trustworthiness—particularly robustness, generalization, and real-time deployment—remain unresolved. This systematic review synthesizes and critically evaluates the existing literature on AI-based protection for inverter-dominated grids, focusing on these three pillars of trustworthiness. We conducted a comprehensive literature search across major databases and preprint repositories, applying inclusion criteria that required original research addressing at least one of the three core themes. After screening titles, abstracts, and full texts, we selected a final set of studies for data extraction and synthesis. The results reveal a rapidly growing but fragmented landscape: a substantial body of work addresses robustness through adversarial training and defense mechanisms, while generalization is less frequently explored, with notable contributions in domain adaptation and zero-shot learning. Real-time deployment emerges as an active theme, with hardware-in-the-loop implementations and latency-aware architectures meeting sub-cycle requirements. However, very few studies simultaneously address all three dimensions, and most optimize for one or two aspects at the expense of others. The absence of standardized benchmarks, adversarial threat models, and real-world validation datasets further hampers cross-study comparability. We conclude that siloed advancements in robustness, generalization, and real-time deployment represent a critical barrier to trustworthy AI in this domain. Future research must prioritize unified frameworks that embed formal robustness guarantees, out-of-distribution detection, and latency-constrained inference within a single protection architecture, alongside open-access testbeds for reproducible evaluation