Data-Centric AI for Zero-Carbon Power Systems Security: A Framework for Learning with Noisy, Sparse, and Heterogeneous Data

Abstract

Zero-carbon grids increasingly rely on pervasive sensing and AI-driven automation, yet most learning engines still assume clean, synchronous data and bolt-on security tools. We introduce a six-layer, data-centric AI framework that (i) raises data quality before inference, (ii) fuses heterogeneous telemetry in real time, and (iii) embeds graph-neural security analytics that adapt to evolving threats. Using four open benchmarks—PSML, PowerGraph, the UCI Smart-Grid Stability set, and GridLAB-D scenarios—we demonstrate: (1) a 55.22 RMSE reconstruction error that preserves trend integrity after severe sparsification; (2) 100 % anomaly-detection accuracy with zero false alarms; and (3) a ≥94 % data-recovery rate plus sub-150 ms response under simultaneous high-load and cyber-attack stress tests. Compared with conventional model-centric pipelines, our architecture eliminates repeated retraining, reduces feature-engineering overhead, and couples defence logic to the same graph topology used for state estimation. The framework therefore offers a scalable blueprint for real-time, secure operation of renewables-dominated grids. We recommend that regulators codify minimum data-quality protocols, operators deploy topology-aware detection models, and software vendors ship AI modules with integrated preprocessing.