Data Driven Prediction Model (DDPM) for Server Inlet Temperature Prediction in Raised-floor Data Centers

Abstract

Data centers account for approx. 1.4% of the world's electricity consumption, of which up to 50% of that power is dedicated to keeping the actual equipment cool. This represents a huge opportunity to reduce data center energy consumption by tackling the cooling system operations with focus on thermal management. This work presents a novel Data Driven Predictive Model (DDPM) for temperature prediction of server inlet temperatures that utilises high resolution empirical temperature measurements from 52 real-life data centers. A knowledge-base of temperature data and related physical features, created via clustering techniques was used to train a series of artificial neural networks (ANN). The ANNs are used to make predictions of server inlet temperatures based on inputs which describe the boundary conditions. The temperature predictions are made for each server rack to estimate the vertical temperature distribution (s-curve) from the bottom to top of the rack spaced at one foot intervals. Each ANN predicts a temperature at a corresponding vertical height for the given inputs, producing the s-curve reconstructed from the combination of ANN outputs. Furthermore, one ANN predicts the s-curve cluster which is used to provide a prediction confidence. The model only requires local boundary conditions such as rack power, perforated tile airflow rate and temperature, ceiling temperature and a rack adjacency identifier (RAI), in addition to average delta T for air conditioners and an identifier for data center layout type. Both RAI and data center identifier are assigned programmatically given a data centers layout information while the rack power, perforated tile airflow rate and delta T can be measured, metered or calculated. Unlike other statistical approaches which are specific to a single data center room, the DDPM is trained with data from a wide range of data centers, therefore can be used to predict server inlet temperatures in many different types of data centers. The prediction is accompanied by a confidence level in the prediction. DDPM results gave a prediction accuracy of 0.76C RMSE with a 0.12 probability of one point on the s-curve crossing the upper or lower bounds of the confidence interval. The model can perform in real-time, giving way to applications for real-time monitoring, input to optimize control of air conditioning units, and can complement sensor networks.