Proper Orthogonal Decomposition-Based Modeling Framework for Improving Spatial Resolution of Measured Temperature Data

Abstract

This paper presents a proper orthogonal decomposition (POD)-based reduced-order modeling framework to improve spatial resolution of measured temperature data in an air-cooled data center. This data-driven approach is applied on transient air temperature data, acquired at the exhaust of a server simulator rack. Temperature data is collected by a distributed thermocouple network at 1 Hz sampling frequency following a step impulse in the rack heat load. The input data are organized in a 2-D array, comprising transient temperature signals measured at various spatial locations. Because its computational time scales logarithmically with the input size, the proposed POD-based approach is potentially useful as an efficient tool for handling large transient data sets. With spatial location being the parameter for the input data matrix, the proposed approach is suitable for rapid synthesis of transient temperature data at new spatial locations. The comparison between POD-based local air temperature predictions and corresponding data indicates a maximum prediction uncertainty of 3.2%, and root mean square prediction uncertainty of 1.9%.