The Impact of Design Space on the Accuracy of Predictive Models in Predicting Chiller Demand Using Short-Term Data

Abstract

Predicting cooling load is essential for many applications such as diagnosing the health of existing chillers, providing better control functionality, and minimizing peak loads. In this study, short-term chiller and total building demand are acquired for five different commercial buildings in the Midwest USA. Four different machine learning models are then used to predict the chiller demand using the total building demand, outdoor weather data, and day/time information. Two data collection scenarios are considered. The first relies upon use of multiple weeks of data collection that includes very warm periods and season transitional periods where the outdoor temperature ranged from very warm to cool conditions in order to envelope all cooling season weather conditions. The second scenario employs use of contiguous data for a several weeks during only the warmest period of the year. The results show that using two or more separate time periods to envelope most of the weather data yields a much more accurate model in comparison to use of data for only one time period. These research findings have importance to energy service companies which often do short term audits (measurements) in order to estimate potential savings from chiller system upgrades (controls or otherwise).