A Bidirectional Gated Recurrent Unit Model for PUE Prediction in Data Centers

Overview

The study develops a BiGRU neural network to predict data centre PUE and compares it with a standard GRU. Training data come from a one-year EnergyPlus simulation of the example data centre model with Singapore weather at 10-minute steps: 52,560 samples with 117 features. The data are therefore simulator output, not metered field data.

Feature selection uses recursive feature elimination with cross-validation (RFECV) with XGBoost, choosing subsets of 21-43 variables covering outdoor conditions, IT power, UPS and CRAC/cooling-coil variables, with hyperparameters tuned for each set. RFECV selected sets of {21, 28, 29, 32, 33, 43} features with the minimum MSE. The 21 features include outdoor dry-bulb, dew-point and wet-bulb temperature, ITE CPU, fan and UPS electric power, air volume flow, CRAC total power, net sensible capacity and SCOP, and cooling-coil power and outlet temperature.

The BiGRU with 21 features gave the lowest error: MAE 0.00038 (MSE 5.6342e-7, R2 0.99673), while the GRU performed best with 28 features (MAE 0.0004181). The paper recalls that Google's neural-network PUE model achieved an MAE of 0.004 +/- 0.005 for a PUE of 1.1, and that Singapore guidance requires new data centres to have a PUE of 1.3 or below at 100 percent IT load.

The key limit is that the data are simulated: the very high accuracy reflects the regularity of the simulator, not field validation. The model is a surrogate of the simulator and should be validated against metered data before being used for guarantees. The document is a preprint, not a peer-reviewed publication. ISO/IEC 30134-2 is named as the related standard.

Key points

1. Dataset: one-year EnergyPlus simulation with 2009 Singapore weather at 10-minute intervals, giving 52,560 samples with 117 features.
2. RFECV selected feature sets of {21, 28, 29, 32, 33, 43} features with the minimum MSE.
3. The 21 features include outdoor dry-bulb, dew-point and wet-bulb temperature, IT CPU, fan and UPS power, air flow, CRAC power, net sensible capacity and SCOP, and coil power and outlet temperature.
4. BiGRU with 21 features achieved the lowest MAE of 0.00038 (MSE 5.6342e-7, R2 0.99673); GRU did best with 28 features (MAE 0.0004181).
5. Cited Singapore guidance requires new data centres to have a PUE of 1.3 or below at 100 percent IT load.
6. Prior work cited: Google's neural-network PUE model achieved an MAE of 0.004 +/- 0.005 for a PUE of 1.1.

Using it in design

• For an operations or digital-twin module, use the feature list as a practical minimum sensor set for PUE prediction: outdoor dry/wet-bulb, IT power, UPS power, CRAC power/capacity/SCOP, coil temperatures and airflow.
• Treat ML PUE models trained on simulation as surrogates of the simulator and validate against metered data before using them for guarantees.
• Keep regional PUE limits (e.g. Singapore 1.3 at full IT load) in the metrics/compliance catalogue.

Who should read it

Operations engineers and data scientists building digital twins and energy forecasting.

Related standards

• ISO/IEC 30134-2

Publication details

Authors

Dhivya Dharshini Kannan; Anupam Trivedi; Dipti Srinivasan

Publisher

arXiv preprint (National University of Singapore)

Year

2025

Document number

arXiv:2512.20161

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.48550/arXiv.2512.20161

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation