Carbon-aware forecasting-enabled scheduling architecture for AI workloads in sustainable distributed systems

Abstract

Artificial intelligence (AI) workloads are rapidly increasing energy demand across distributed cloud and edge computing infrastructures. Existing carbon-aware scheduling studies often focus on optimisation strategies, with limited integration between workload telemetry, forecasting, and deployable scheduling architectures. This paper presents real AI workload profiling and near-term carbon-intensity forecasting as the empirical foundations for a carbon-aware scheduling architecture, organised around an Observe–Reason–Act–Learn (ORAL) operational pipeline inspired by adaptive feedback-loop systems. The study focuses on AI workload profiling, near-term carbon-intensity forecasting, and architecture-level scheduling support for energy- and carbon-aware distributed systems. A workload profiling study is conducted using GPU, CPU, memory, power, and workload-level carbon attribution metrics across representative AI workloads. In addition, near-term forecasting is evaluated using Ridge Regression, XGBoost, and Random Forest models with Singapore-zone carbon-intensity data across 1h–4h horizons. Forecasting achieves high predictive accuracy, with tree-based models attaining mean absolute error below 0.7 gCO₂/kWh and improving over a persistence baseline by up to 96.4%, while all models achieve R² above 0.98, indicating that near-term carbon intensity can be estimated reliably across these horizons. The Act-phase scheduling decision and the Learn-phase feedback mechanism are specified concretely with their closed-loop experimental evaluation identified as ongoing work. Overall, the study contributes workload profiling evidence, preliminary forecasting validation, and architectural design considerations toward practical energy- and carbon-aware distributed systems for AI workloads.