Hybrid Modelling of a Two-Stage Refrigeration System

Refrigeration cycles employed for process cooling in industrial settings are known to be very energy-intensive. With the rise in energy demand across the globe, it is imperative that these systems are monitored and optimised to improve thermal efficiency and minimise energy consumption. While process simulators have been widely employed for process monitoring of refrigeration systems for optimisation and "what-if" scenarios, it is generally not suitable for remote monitoring due to their high computational cost. This is where machine learning (ML) models stand out, as they can be executed up to several orders of magnitude quicker than their corresponding first principle simulation model. In this paper, a case study is presented for a two-stage refrigeration system that covers data management, hybrid modelling (combining process simulation and ML), optimisation and online deployment. This case study demonstrates the workflow for constructing a refrigeration system ML model and deploying it to an online monitoring and prediction dashboard. Five different ML algorithms, namely multivariate linear regression (MLR), regression trees (RTs), random forests (RFs), extreme gradient boosting (XGBoost) and artificial neural networks (ANNs) were tested and evaluated to determine the most suitable algorithm to represent the behaviour of the refrigeration system. The results have shown that the XGBoost algorithm outperformed the other four algorithms as demonstrated by the lowest mean absolute error (MAE) and root mean squared error (RMSE). A significant improvement in performance was achieved after the XGBoost model was optimised via hyperparameter tuning and displayed an 86.40% and 82.35% improvement in terms of the MAE and RMSE. The results obtained from this study serve as a good benchmark for future modelling of refrigeration systems with XGBoost.