Domain-invariant representation learning for generalizable chiller model: a real-world case study

Abstract

Data-driven models have been widely adopted due to their ease of modeling and good prediction accuracy. However, they face a challenge in ensuring predictive performance on unseen datasets, especially when the training dataset is imbalanced. For this reason, Transfer learning (TL), that leverages source domain datasets, has attracted attention. However, TL is often limited by its one-way learning process, which transfers knowledge from a data-rich (source) domain to a data-scarce (target) domain. In this regard, this study proposes a Domain-Invariant Representation Learning (DIRL) modeling approach that extracts generalizable knowledge through a bidirectional learning framework. To realize it, the authors develop a DIRL chiller model using two real-life chillers’ datasets where hidden layers are shared within an artificial neural network (ANN). Relevant data were collected at the sampling time of one hour between April 2020 and December 2023.

For comparison, the following five simulation models of the two-chillers were cross-compared in terms of model accuracy and extrapolation ability: (1) an individual ANN model, (2) an ANN model developed by combined data from two chillers, (3) a transfer learning model developed by the other chiller data, (4) a transfer learning model developed by a physics-based model, and (5) a DIRL model. The predictive performance of all five models was satisfactory for the target chiller by achieving a mean average error (MAE) = 0.41–0.49 and a coefficient of the variation of the root mean square error (cvRMSE) = 7.0–8.3 % for the coefficient of performance (COP). In contrast, the combined-data ANN and DIRL presented superior predictive performance by achieving an MAE = 0.10–0.13 and a cvRMSE = 2.0–3.1 %. The DIRL model demonstrated best superior extrapolation ability with an MAE = 0.36 and a cvRMSE = 8.5 %. As a result, the DIRL model achieved improvements of 0.81 of MAE and 16.4 % of cvRMSE for chiller COP prediction, and 0.83 in MAE and 17.4 % in cvRMSE for extrapolation ability, compared to the individual ANN model. By leveraging bidirectional learning with combined datasets and a shared feature extractor, the DIRL chiller model can infer general chiller knowledge while maintaining a consistent predictive performance in terms of both accuracy and extrapolation ability.