Variable importance for chiller system optimization and sustainability

Research output: Contribution to journalArticlepeer-review


Very few studies have considered modelling the statuses of loading and component combinations (LC) in chiller systems. This study uses a decision tree (DT) and a neural network (NN) to model accurately six output levels of LC in terms of 11 predictor variables. The NN model gave a higher correct prediction of 89.52% than 84.44% in the DT model. A variable dropout analysis from the DT and NN models generalizes the top five significant variables: the statuses of cooling towers and primary chilled water pumps, and the supply temperature, return temperature and flow rate of chilled water. The system coefficient of performance (SCOP) modelled by NN has an improved R 2 of 84.18% versus 74.62% when the significant variables are included as inputs. The SCOP maximization brings three optimal strategies—implementing chiller sequencing, operating components in pairs and optimizing chilled water supply temperature—which reduce CO2 emissions by 99,815–346,987 kg.

Original languageEnglish
Pages (from-to)504-523
Number of pages20
JournalEngineering Optimization
Issue number3
Publication statusPublished - 2022


  • Cooling tower
  • decision tree
  • neural network
  • water-cooled chiller


Dive into the research topics of 'Variable importance for chiller system optimization and sustainability'. Together they form a unique fingerprint.

Cite this