Sustainable Grid Integration of AI Data Centers: Carbon-Aware Co-Optimization of Training Workloads and Battery Storage Under Renewable Uncertainty

Overview

The study asks whether carbon-aware computing is audited at system level: workloads follow scalar price or carbon signals, but claimed savings need not survive network-wide dispatch. The authors develop a two-stage, risk-averse framework that schedules checkpoint intervals, workload migration, dynamic voltage and frequency scaling (DVFS) and behind-the-meter storage jointly with unit commitment, pricing system CO2 beyond the market-internalised level. IEEE 39- and 118-bus systems are mapped to 2025 French and German measurements.

Each data centre's BESS is modelled as a 2-hour system sized at 0.38 kW per GPU (a checkpoint-driven dip-masking rule), with constant round-trip efficiency, daily energy neutrality and SOC fixed at 0.5 at the start and end of the day. Carbon-priced dispatch cut system CO2 by 29.5% at a generation-cost premium below 0.01%; more than 99% of the cut came from generation redispatch, an operator-side instrument.

AI-load flexibility adds 0.23% of system emissions today, about one quarter of the data-centre fleet's own footprint (the modelled fleet drew about 627 MWh/day, about 152 t/day CO2; flexibility abated 39 t/day, about 26%). At gigawatt scale it rises to 3.3% and cuts mean worst-day load shedding from 849 to 431 MWh (-49%). Storage-only flexibility cut emissions by only 0.02% beyond the carbon signal, against 0.22% for migration plus DVFS; on the 118-bus case storage alone was 0.3 t/day worse than the inflexible reference because of round-trip losses.

Following an average carbon-intensity signal can raise system emissions even compared with carbon-blind co-optimised dispatch. Limits: the zonal mapping is a reduced form that keeps network physics but not the true European topology, so absolute magnitudes carry error but directions and rankings do not; average rather than marginal emissions are used; day-ahead prices are assumed known and common across zones; and the thermal model uses static PUE and storage a constant efficiency.

Key points

1. Each data centre's BESS is a 2-hour system sized at 0.38 kW per GPU, with constant round-trip efficiency, daily energy neutrality and SOC 0.5 at the start and end of the day.
2. Carbon-priced dispatch cut system CO2 by 29.5% at a generation-cost premium below 0.01%; more than 99% of the cut came from generation redispatch.
3. Storage-only flexibility (case B3) cut emissions by only 0.02% beyond the carbon signal, against 0.22% for migration plus DVFS; on the 118-bus case storage alone was 0.3 t/day worse than the inflexible reference.
4. The modelled AI fleet drew about 627 MWh/day (about 152 t/day CO2); flexibility abated 39 t/day, about 26% of the fleet's own emissions.
5. At gigawatt scale, AI-load flexibility cut mean worst-day load shedding from 849 to 431 MWh (-49%) and gave a 3.3% system emission reduction.
6. Scheduling against the scalar average carbon-intensity signal can raise system-level emissions, even compared with carbon-blind co-optimised dispatch.

Using it in design

• Do not size behind-the-meter battery storage on an emission-reduction case; round-trip losses mean storage alone adds nothing and can raise emissions. Justify it with a concrete duty such as masking checkpoint dips.
• Present AI-training checkpoint-interval, migration and DVFS flexibility in grid-connection negotiation as a capacity and load-shedding mitigation asset.
• Do not use the average carbon-intensity signal alone in carbon-aware scheduling; verify the effect on system-level dispatch.

Who should read it

Designers and energy strategists planning grid impact, flexibility and storage strategy for AI campuses.

Document contents

1. 1. Introduction
2. 2. Materials and Methods (2.1 System Overview and Modeling Assumptions, 2.2 Checkpoint-Shaped AI Training Load, 2.3 Data-Center Migration, DVFS, and Service Quality, 2.5 Zonal Carbon and Price Signals from Public Data, 2.6 Network and Generation Model, 2.7 Two-Stage Risk-Averse Formulation)
3. 3. Experimental Design, Baselines, and Metrics (3.1 Baselines, 3.2 Metrics, 3.3 Test Systems and DC Layouts, 3.4 Computational Setup)
4. 4. Results (4.1 Co-Optimization Against Eight Baselines, 4.2 Flexibility Analysis on Representative Day Operation, 4.4 Ablation Studies, 4.6 Scale Effects, 4.7 Sensitivity to Structural Model Choices)
5. 5. Discussion (5.1 Principal Findings, 5.2 Interpretation, 5.3 Sustainability Implications, 5.4 Limitations)
6. 6. Conclusions

Related standards

• IEEE 1547
• IEC 62933

Read the document

Publication details

Authors

Junqing Zhang; Yi Wang

Publisher

Sustainability (MDPI AG)

Year

2026

Volume

18

Issue

18

Pages

9504

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.3390/su18189504

Licence

CC BY 4.0 — share with attribution

Licence evidence

www.mdpi.com

Citation