Carbon-Aware AI Computing: Optimizing Data Center Power Consumption Using Workload, Environmental, and Grid Emissions Data

This research proposes a carbon-aware AI computing framework that integrates workload telemetry, real-time environmental conditions and grid carbon intensity to optimise power consumption and reduce the carbon footprint of AI workloads in high-performance data centers. Drawing on information theory and sustainable computing, it uses a quantitative design combining the MIT Supercloud job dataset, Weather Underground historical weather data and WattTime grid emissions data across three phases: statistical analysis, predictive machine learning modelling, and simulation of carbon-aware scheduling policies. The study aims to help data center operators anticipate energy use, manage costs and shift workloads to periods of cleaner energy.