Environment data processing for a data centre (2)

Abstract

Data centres form the backbone and brains behind contemporary digital services. But in order for data centers to function and be reliable, a significant amount of energy is used for the purpose of cooling servers. This project aims to develop predictive models to accurately forecast the cooling power consumption required to cool data centres, using the TDC2.0 dataset from an 11-month experiment conducted in a Singapore air-cooled data centre testbed. By leveraging machine learning techniques, such as tree-based models, neural networks and foundation models, robust models were created that can model cooling power consumption based on environmental and operational parameters. From the experiments, it was found that just using the temporal dependencies and autocorrelation properties of the cooling power consumption was sufficient to give accurate predictions, outperforming models which incorporated environmental factors and setpoint information as the features. It was also found that zero-shot forecasting performance of the time-series foundation model TimesFM v1.0 and v2.0 compared to the other models used was poor, possibly due to the limited context length.

Additionally, using feature importance metrics and recursive feature elimination from LightGBM and XGBoost, the most important environmental and setpoint features that influenced cooling power consumption were explored in-depth and analyzed to better optimize for cooling power efficiency.