Artificial Intelligence-Based Temperature Twinning and Pre-Control for Data Center Airflow Organization

Overview

The study asks how AI-based temperature forecasting and air-conditioning control can prevent hotspots before they form and improve energy efficiency. The test room is 5.4 m high with a 1.0 m raised floor, five contained micro-modules of 40 cabinets (200 total, 2.2 m high, 5 kW each) and 70% open floor tiles. Fourteen 100 kW units (5 duty + 2 standby per side) of 25,000 m³/h each sit in side AC rooms. The room has 728 sensor points sampled every 15 min, with temperature accuracy ±0.5 °C over -10 to +50 °C.

The LSTM forecasts supply, return, cold-aisle and hot-aisle temperatures from the temperature and equipment power of the past 15 time steps. A DQN controller then adjusts air-conditioning output so that predicted hotspots do not occur. Setpoints were supply air 26 ± 1 °C, return air 39 ± 1 °C and cold aisle 27 ± 1 °C, with a target prediction error of no more than ±0.4 °C.

Results: LSTM mean absolute errors were 0.32 °C for supply air, 0.21 °C for return air, 0.36 °C for the cold aisle and 0.19 °C for the hot aisle, with cabinet power predicted within ±2 kW. The DQN controller cut PUE by an average of 2.6% and WUE by 2.5% over a year.

Limits: the object is limited because the building structure could not be modified, the cooling and air-supply type is simple, and only particular AI verification was done. As a learned, sensor-based model rather than a CFD surrogate, it shows what instrumentation operational twins need.

Key points

1. Test room: 5.4 m height, 1.0 m raised floor, five contained micro-modules of 40 cabinets (200 total, 5 kW each, 2.2 m high), 70% open floor tiles; fourteen 100 kW units (5 duty + 2 standby per side) of 25,000 m³/h each in side AC rooms.
2. The room has 728 sensor points sampled every 15 min, with temperature accuracy ±0.5 °C over -10 to +50 °C.
3. Setpoints: supply air 26 ± 1 °C, return air 39 ± 1 °C and cold aisle 27 ± 1 °C, with a target prediction error of no more than ±0.4 °C.
4. LSTM mean absolute errors: 0.32 °C supply air, 0.21 °C return air, 0.36 °C cold aisle and 0.19 °C hot aisle; cabinet power predicted within ±2 kW.
5. The DQN controller cut PUE by an average of 2.6% and WUE by 2.5% over a year.

Using it in design

• For an operational thermal twin, define adequate sensor density covering supply, return, cold-aisle and hot-aisle temperatures and equipment power (728 points at 15-minute sampling in this study).
• Set the prediction error target (±0.4 °C) and setpoints (supply 26 °C, return 39 °C, cold aisle 27 °C) as twin validation criteria.
• Design air-conditioning control as pre-control that acts before hotspots form, and weigh the PUE/WUE gain (about 2.5%) against sensor and control investment.

Who should read it

Engineers and researchers developing data-centre cooling control and operational twins.

Document contents

1. 1. Introduction
2. 2. Materials and Methods
3. 3. Digital Twin (Simulation) Module (3.1 Application of Digital Twin Technology in Data Center Cooling Systems, 3.2 Time Series Based Temperature Prediction Model)
4. 4. Optimal Control Module: Hotspot and Fault Pre-Control and Group Control (4.1 Hot Spots and Fault Pre-Control, 4.2 Optimised Control Module for Group Control)
5. 5. Experiments (5.1 Data Center Simulation Platform, 5.2 Experimental Data, 5.3 Experimental Setup, 5.4 Experiments and Analysis of Data Center Task Prediction)
6. 6. Conclusions

Related standards

• ASHRAE TC 9.9 Thermal Guidelines for Data Processing Environments
• ISO/IEC 30134-2 (PUE)
• ISO/IEC 30134-9 (WUE)

Read the document

Publication details

Authors

Na Huang; Xiang Li; Quanming Xu; Ronghao Chen; Huidong Chen; Aidong Chen

Publisher

Energies (MDPI AG)

Year

2023

Volume

16

Issue

16

Pages

6063

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.3390/en16166063

Licence

CC BY 4.0 — share with attribution

Licence evidence

www.mdpi.com

Citation