Grid Accommodation of Dynamic HPC Demand (EE HPC State of the Practice Workshop, 2019)

Overview

This document is a paper from the Energy Efficient HPC Working Group (EE HPC WG) for its State of the Practice Workshop in Kyoto in 2019; the authors are Grant L. Stewart, Gregory A. Koenig, Jingjing Liu, Anders Clausen, Sonja Klingert and Natalie Bates. It explains why rapid power swings from large supercomputers can disturb voltage on the local grid and what can mitigate them.

The paper reviews the basics of voltage fluctuation, reactance and grid stiffness, lists stabilising devices, and presents four anonymised HPC site case studies built from questionnaires and interviews. The sites differ in transmission voltage, nearby generation and VAR support. Stabilising devices include generators with automatic voltage regulation, capacitor banks, SVCs and STATCOMs. The scale is notable: exascale systems are expected to have 30-40 MW peak demand and up to 20 MW load swings in 15 ms (one 60 Hz cycle) to several seconds. One cited deployment saw 1.9 MW fluctuations causing 9% voltage fluctuation during initial testing; another study measured about 0.4% for 1.5 MW, anticipating up to 4%. ORNL's Summit has a 13 MW peak demand.

The case-study sites range from a developing site (1.3 MW peak with 1.0 MW swings, planning 14 MW) to sites on 345 kV and 500 kV networks with heavy VAR support. The authors conclude that dynamic-load modelling with the utility is needed early in planning, because mitigation measures can be comparable in cost to the computing system itself. These figures show that the effect depends on the grid's strength at the point of connection as much as on the size of the swing.

Key points

1. Exascale systems are expected to have 30-40 MW peak demand and up to 20 MW load swings in 15 ms (one 60 Hz cycle) to several seconds.
2. One deployment saw 1.9 MW fluctuations causing 9% voltage fluctuation during initial testing; another study measured about 0.4% for 1.5 MW, anticipating up to 4%.
3. ORNL's Summit has a 13 MW peak demand.
4. Mitigating devices: generators with automatic voltage regulation, capacitor banks, SVCs and STATCOMs.
5. Case-study sites range from a developing site (1.3 MW peak, 1.0 MW swings, planning 14 MW) to sites on 345 kV and 500 kV networks with heavy VAR support.
6. Mitigation can be comparable in cost to the computing system itself, so early planning with the utility is needed.
7. Related standard: IEEE 1453 (voltage fluctuations/flicker).

Using it in design

• Site/grid-connection screens should capture expected load-step magnitude and ramp time and short-circuit strength at the point of connection.
• Warn when large fast load swings are planned on a weak grid without VAR compensation or energy-storage buffering.
• Do dynamic-load modelling with the utility at the start of planning.
• Add the cost of stabilising devices (SVC, STATCOM, capacitor banks) to the project budget early.

Who should read it

Electrical designers and grid-connection engineers planning sites with large, fast-changing HPC and AI loads.

Related standards

• IEEE 1453 (voltage fluctuations/flicker)

Publication details

Authors

Grant L. Stewart; Gregory A. Koenig; Jingjing Liu; Anders Clausen; Sonja Klingert; Natalie Bates

Publisher

Energy Efficient HPC Working Group

Year

2019

Document number

EE HPC SOP Workshop, Kyoto, 2019

Event

EE HPC State of the Practice Workshop 2019, Kyoto

Document language

English

Source type

Industry body

Access

Free to read

Licence

Copyright retained; citation only — the publisher’s terms apply to the original document

Citation