A Comparison of the Performance of a Data-Driven Surrogate Model of a Dehumidifier with Mathematical Model of Humidification-Dehumidification System

Abstract

In this study, we investigate the potential use of machine learning regression to develop a surrogate trained model based on experimental data for the dehumidifier component of a humidification-dehumidification desalination system. We study different machine learning regression approaches such as linear regression, Gaussian process regression, neural network and support vector regression to compare their performance. Furthermore, hyper-parameter optimization was carried out for the neural network to find the optimized architecture. We find that the mathematical model has an average uncertainty of 10 percent, while the surrogate model based on support vector regression had an average uncertainty of 7 percent in its prediction. In addition, Gaussian process regression required a large amount of computational effort, while the uncertainty for the surrogate model is higher than mathematical model. Hyper-parameter optimization led to a surrogate model with an accuracy of 99 percent, which could help investigate the system performance under different conditions.