How does the control logic influence the establishment of a data-driven chiller model?

Abstract

The modelling of chiller performance is critical for chiller optimal control, and fault detection and diagnosis (FDD). Different kinds of chiller models including sophisticated mechanistic models (white-box models), purely data-driven models (black-box models like artificial neural networks, ANN), and semi-physical models (grey-box models like empirical equations) have been proposed and tested. Due to the development of machine learning techniques, data-driven models have become more popular recently. The performance of the established data-driven model (accuracy, robustness, generalization) could significantly affect the model application. To enhance the model performance, a lot of studies have been carried out on investigating and modifying the model structure. However, the influence of the data quality on model training has not been sufficiently studied. When adopting historical data to train models, the data distribution is highly correlated to the control logic. So how does the control logic influence the establishment of data-driven chiller models by affecting the data distribution? In this study, experiments are conducted on an air-cooled chiller under model-free stochastic control to acquire rich and variable operational dataset; then the dataset is grouped into three corresponding to different chiller control logics. Finally, three models trained by three training datasets are evaluated, and the results suggest that when establishing data-driven chiller models, preliminary stochastic operation is cost-effective to acquire rich data for robust chiller modelling.