AbstractRecent advances in Artificial Intelligence (AI) and Deep Reinforcement Learning (DRL) have enabled the development of intelligent and adaptive decision-making systems capable of operating in highly dynamic and adversarial environments. Traditional rule-based and shallow machine learning techniques lack the capability to model complex temporal dependencies and hidden attack patterns associated with sophisticated cyber-attacks such as False Data Injection (FDI) and Denial-of-Service (DoS/DDoS) in Smart grid Cyber-Physical Systems (CPS). The framework developed in this study was built around a fundamentally different philosophy - a Constrained Proximal Policy Optimization (CPPO) agent with a Control-Barrier Shield (CBS) to deliver autonomous, safety-aware cyber defense for smart grid environments. The agent learns a repertoire of protective responses — including node isolation, communication path reconfiguration, and controlled load redistribution — while the CBS component enforces hard physical constraints on voltage, frequency, and thermal limits before any action reaches the grid. This design keeps safety enforcement structurally separate from the learning process, making the overall system both interpretable and reliable. Spatio-temporal awareness is provided through a convolutional–recurrent neural network (CRNN) with an attention mechanism that processes multivariate grid telemetry across both node space and time. This perceptual layer feeds cleaned, contextually rich representations into the CPPO agent, maintaining decision quality even when incoming measurements carry adversarial distortions. Validation across experiments on IEEE 39-bus and 118-bus benchmark systems demonstrate 36% faster attack recovery, 31% improvement in operational stability, and 62% fewer safety violations against competitive baselines. The framework also generalizes effectively to previously unseen attack configurations, consistently holding system variables within safe operating boundaries. Proposed system outperformed benchmark IEEE 39-bus and 118-bus systems, 36% faster attack recovery, 31% improvement in operational stability, and 62% fewer safety violations against competitive baseline systems. The framework also generalized effectively to unseen attack configurations and consistently holding system variables within the safe operating boundaries. Keywords: Deep Reinforcement Learning, CPPO (Constrained Proximal Policy Optimization), CRNN(convolutional–recurrent neural network), Control-Barrier Shield, Spatio-Temporal Detection, Adaptive Cyber Defense, Smart Grid Security, Cyber-Physical Systems, FDI ( False Data Injection ), Denial-of-Service ( DoS/DDoS).
Kumar et al. (Wed,) studied this question.