Accuracy and Robustness Against Covariate Shift of Water Chiller Models

The optimal energy management of multiple chiller systems calls for the construction of mathematical models of chiller energy efficiency. The existing grey- or black-box models include parameters that have to be estimated from experimental data. So far, the predictive capabilities of alternative models have been assessed and compared on data sets created by laboratory tests or provided by chiller manufacturers. In an Industry 4.0 context, the continuous monitoring and collection of field data discloses new opportunities but raises also robustness issues that are herein addressed. Herein, exploiting an extensive experimental dataset collected over a six-month period, four literature models and a new machine learning approach are compared. The second objective is assessing the robustness of the five models against covariate shifts, i.e. variations in the statistical distribution of the input variables that occur across different months. The grey-box Gordon-Ng model, though less accurate in nominal conditions than the Bi-quadratic and Multivariate polynomial models, proves however more robust against covariate shifts. The best performances, both in term of accuracy and robustness, are however provided by the two machine learning methods, with the Gaussian Process model performing better than the MLP artificial neural network.