Discovering a robust machine learning model for predicting the productivity of a solar-driven humidification-dehumidification system

This work introduces a complete study of freshwater productivity prediction of a solar-driven humidification-dehumidification unit (HDH) based on experimental and machine learning methods. Freshwater productivity besides other operational variables was first measured under a series of twenty outdoor experiments each lasted for 4 h. According to these experiments, average accumulated productivity reached up to 10.8 L/m². Furthermore, these recorded data were used to construct machine learning models for predicting the hourly freshwater productivity, cost, and GOR of the HDH system. Four types of machine learning algorithms were constructed including artificial neural network, random forest, linear support vector machine, and support vector machine. More importantly, the hyperparameters of these algorithms were optimized based on Bayesian optimization algorithm (BOA). Measured variables of solar radiation, meteorological conditions, carrier air flow rate, and temperatures of air and water paths were used as inputs for prediction models. The important feature of these descriptors on output was also estimated and presented based on the trained random forest (RF) model. As a comparsion, the artificial neural network (ANN) and RF model can achieve a more accurate prediction of the system hourly productivity, cost, and GOR than the other models, where these values of MSE and R² reached (0.0999, and 0.975) and (0.088, and 0.977), respectively. Accordingly, the ANN-BOA model can provide benefits for modeling of hourly productivity and RF-BOA can provide accurate strategies for optimizing the performance of the solar-driven HDH unit.