Native AI-based Predictive Operational Resiliency in Cyber-Physical Energy Systems

Operational resiliency becomes a key requirement to prevent critical recall of distributed energy resources (DER) in Cyber-Physical Energy Systems (CPES). Thus, CPES requires methods and metrics to observe and understand the potential faults of unwanted cyber-physical events. This paper introduces a new Native-AI-driven predictive maintenance framework to ensure the operational resiliency of CPES. First, a system model is designed to capture operational states such as normal, service mode, and faults by observing the operational behavior of DERs. Second, a predictive maintenance framework is proposed and developed by introducing a dual-method feature selection mechanism while the features of critical recall states are selected by tailored Gini and permutation-based importance metrics. Third, the predictive model is designed by leveraging tree-based models that utilize cyclical temporal encoding and lag features, while a Long Short-Term Memory (LSTM) family models employ sequence learning with normalized temporal and angular representations. Finally, the proposed framework can achieve up to 99.97% Critical Recall detection accuracy with an average of 82.5%. As a result, this work advances AI-driven predictive maintenance by reducing downtime by 15–30% in simulated scenarios.