Green ICT and Sustainable Computing: Evaluating Carbon Footprint Metrics in Cloud- to-Edge AI Workloads

The exponential expansion of Artificial Intelligence (AI) model training
and inference workloads across cloud-to-edge infrastructures has led to
staggering energy consumption and operational carbon emissions. While
cloud datacenters and industrial edge nodes leverage power usage
effectiveness (PUE) metrics, traditional workload scheduling algorithms
remain carbon-oblivious, optimizing solely for execution cost or
completion latency. This paper introduces the Dynamic Carbon-Aware
AI Workload Scheduler (DCAS), a mathematical optimization model
designed to minimize operational carbon footprints (gCO2eq) across
geo-distributed cloud datacenters and industrial edge processing
clusters. DCAS formulates a multi-objective Mixed-Integer Linear
Programming (MILP) model that dynamically shifts compute-heavy
Deep Learning (DL) training and batch inference jobs in space (spatial
routing to regions with cleaner energy grids) and time (temporal deferral
to match peak solar and wind availability). Evaluating 1,000 deep
learning training jobs across 10 geo-distributed cloud and edge facility
nodes over a 30-day simulation window using real-time grid carbon
intensity feeds, empirical results show that DCAS reduces total

operational carbon emissions by 58.4% compared to static cost-
optimized baselines, while restricting average Service Level Agreement

(SLA) delay penalties to less than 4.2%. Furthermore, spatial-temporal
workload shifting paired with edge battery energy storage systems
(BESS) reduces peak grid power demand by 42.1%, delivering a scalable
blueprint for sustainable Green ICT infrastructure in Industry 4.0
environments.