Harnessing the power of neural networks for the investigation of solar-driven membrane distillation systems under the dynamic operation mode
School of Engineering
Accurate modeling of solar-driven direct contact membrane distillation systems (DCMD) can enhance the commercialization of these promising systems. However, the existing dynamic mathematical models for predicting the performance of these systems are complex and computationally expensive. This is due to the intermittent nature of solar energy and complex heat/mass transfer of different components of solar-driven DCMD systems (solar collectors, MD modules and storage tanks). This study applies a machine learning-based approach to model the dynamic nature of a solar-driven DCMD system for the first time. A small-scale rig was designed and fabricated to experimentally assess the performance of the system over 20 days. The predictive capabilities of two neural network models: multilayer perceptron (MLP) and long short-term memory (LSTM) were then comprehensively examined to predict the permeate flux, efficiency and gain-output-ratio (GOR). The results showed that both models can efficiently predict the dynamic performance of solar-driven DCMD systems, where MLP outperformed the LSTM model overall, especially in the prediction of efficiency. Additionally, it was indicated that the accuracy of the models for the prediction of GOR can be significantly improved by increasing the size of the dataset.