Design and Prototype of a Microgrid for Enhancing Reliability of Power Supply for Artificial Intelligence Data Centers

Overview

The study asks whether on-site solar and batteries can improve power continuity for an AI data centre exposed to utility outages, and how such a system is sized. The case is a 5.85 MW peak data centre in Cheyenne, Wyoming. The microgrid is designed so that solar PV covers 60% of demand and the utility covers the remaining 40%, with a battery storage system that supplies critical loads at night and during outages.

The method is deliberately simple. Daily energy use is derived from a scaled 24-hour load curve, PV output comes from NREL production estimates, and plain assumptions are used for inverter and battery losses and depth of discharge. Sizing follows an energy-balance chain: daytime and night load split, conversion losses, usable depth of discharge and power rating against the night peak. A small 12.8 V prototype then shows a microcontroller managing critical and non-critical loads according to battery state of charge (SOC).

The results are consistent at design level: a PV array producing about 82,367 kWh/day meets the 81,857 kWh/day target, and the battery is sized at 50,558 kWh and 2,775 kW. In prototype tests all loads run at 90% SOC and non-critical loads are shed automatically as SOC falls. The authors present this as evidence that the microgrid improves reliability.

The limits are clear. The reliability benefit is qualitative, with no formal availability or loss-of-load calculation. Sizing is by hand, and outage scenarios, runs of cloudy days, battery ageing and cost are out of scope. A 12.8 V prototype does not exercise power electronics, protection or MW-scale integration. The work is useful as a conceptual pre-sizing example, not as design validation.

Key points

1. The case-study site has 5.85 MW peak demand and about 5.35 MW critical load, of which servers are 3.5830 MW and cooling 0.8845 MW.
2. Covering 60% of demand needs 81,857 kWh/day from PV after 5% DC-AC and 10% battery losses.
3. A 13,600 kW two-axis tracking PV array (14.08% system losses, 7.76 kWh/m²/day) yields about 82,367 kWh/day.
4. The battery is sized at 2,775 kW (60% of a 4,625 kW night peak) and 50,558 kWh, giving 18.2 h at full discharge or 12.8 h when 70% of capacity is discharged.
5. Prototype control sheds non-critical loads and isolates the battery for charging when SOC falls below 30%; all loads run at 90% SOC.
6. With 60% PV and 40% utility, the design also aims to cut carbon footprint and utility draw.
7. The study contains no formal reliability calculation; the benefit is argued qualitatively.

Using it in design

• Follow the energy-balance chain in a PV/BESS sizing helper (daily load split, conversion losses, usable depth of discharge, power rating against night peak); report usable-energy duration rather than nameplate duration and make DoD an explicit input.
• Back reliability claims for on-site renewables with an availability or loss-of-load calculation, and do not credit PV as firm backup without storage sized for the outage duration.
• Offer SOC-based load shedding (critical versus non-critical) as a configurable control rule in the backup-power model.

Who should read it

Engineers and students doing first-pass on-site solar and battery sizing for data-centre sites.

Document contents

1. 1 Introduction
2. 2 Theoretical Design of Microgrid (data center power consumption analysis, solar PV, sizing of battery energy storage system, inverter size)
3. 3 Prototype Design and Parameters
4. 4 Prototype (implementation of prototype, evaluation of prototype performance)
5. 5 Lessons Learned
6. 6 Conclusion

Read the document

Publication details

Authors

Ramon Zintzun; Ethan Hoang; Jared Wong; Akanksha Katta; Vinh Nguyen; Alex Villaneda; Gerardo Maximo; Ha Thu Le

Publisher

WSEAS TRANSACTIONS ON POWER SYSTEMS (World Scientific and Engineering Academy and Society (WSEAS))

Year

2026

Volume

21

Pages

106

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.37394/232016.2026.21.10

Licence

CC BY 4.0 — share with attribution

Licence evidence

doi.org

Citation