Chiller system performance management with market basket analysis

Abstract

Purpose This study aims to apply association rule mining (ARM) to uncover specific associations between operating components of a chiller system and improve its coefficient of performance (COP), hence reducing the electricity use of buildings with central air conditioning. Design/methodology/approach First, 13 operating variables were identified, comprising measures of temperatures and flow rates of system components and their switching statuses. The variables were grouped into four bins before carrying out ARM. Strong rules were produced to associate the variables and switching statuses with different COP classes. Findings The strong rules explain existing constraints on practising chiller sequencing and prioritise variables for optimisation. Based on strong rules for the highest COP class, the optimal operating strategy involves rescheduling chillers and their associated components in pairs during a high load operation. Resetting the chilled water supply temperature is the next best strategy, followed by resetting the condenser water entering temperature, subject to operating constraints. Research limitations/implications This study considers the even frequency method with four bins only. Replication work can be done with other discretisation methods and different numbers of classes to compare potential differences in the bin ranges of the optimised variables. Practical implications The strong rules identified by ARM highlight associations between variables and high or low COPs. This supports the selection of critical variables and the operating status of system components to maximise the COP. Tailor-made optimisation strategies and the associated electricity savings can be further evaluated. Originality/value Previous studies applied ARM for chiller fault detection but without considering system performance under the interaction of different components. The novelty of this study is its demonstration of ARM’s intelligence at discovering associations in past operating data. This enables the identification of tailor-made energy management opportunities, which are essential for all engineering systems. ARM is free from the prediction errors of typical regression and black-box models.