Selecting the model is an important and essential stepin model based fault detection and diagnosis (FDD).Factors that are considered in evaluating a modelinclude accuracy, training data requirements,calibration effort, generality, and computationalrequirements. The objective of this study was toevaluate different modeling approaches for theirapplicability to model based FDD of vaporcompression chillers.Three different models were studied: the Gordon andNg Universal Chiller model (2nd generation) and amodified version of the ASHRAE Primary Toolkitmodel, which are both based on first principles, andthe DOE-2 chiller model, as implemented inCoolToolsTM, which is empirical. The models werecompared in terms of their ability to reproduce theobserved performance of an older, centrifugal chilleroperating in a commercial office building and anewer centrifugal chiller in a laboratory.All three models displayed similar levels of accuracy.Of the first principles models, the Gordon-Ng modelhas the advantage of being linear in the parameters,which allows more robust parameter estimationmethods to be used and facilitates estimation of theuncertainty in the parameter values. The ASHRAEToolkit Model may have advantages when refrigeranttemperature measurements are also available. TheDOE-2 model can be expected to have advantageswhen very limited data are available to calibrate themodel, as long as one of the previously identifiedmodels in the CoolTools library matches theperformance of the chiller in question.
展开▼
