Federated Learning for Power Cyber-Physical Systems: Toward Secure, Resilient, and Explainable Intelligence

The digital transformation of power cyber-physical systems (CPSs) introduces unprecedented opportunities for optimization, forecasting, and real-time control, while simultaneously exposing critical vulnerabilities in data security, system resilience, and operator trust. Federated Learning (FL) provides a promising paradigm by enabling collaborative intelligence without raw data sharing, yet traditional approaches fall short in safety-critical energy infrastructures. This review advances the state of the art by presenting a holistic perspective on secure, resilient, and explainable FL for Power CPSs. We first analyze emerging threats—including model poisoning, backdoor insertion, and cross-layer false data injection—and map them to existing defenses such as robust aggregation, Byzantine resilience, differential privacy, and zero-trust authentication. We then synthesize architectural innovations, including personalized FL, digital twin–enhanced validation, and human-in-the-loop trust calibration, highlighting their potential to address system heterogeneity and operational risks. Real-world applications in load forecasting, intrusion detection, EV coordination, and microgrid control are surveyed to demonstrate feasibility. Finally, we outline future research directions linking adversarial robustness, explainability, scalable integration, and governance frameworks. This work positions federated learning as a cornerstone for trustworthy intelligence in next-generation power systems.