A holistic framework for improving the prediction of reverse osmosis membrane performance using machine learning

Abstract

Accurate prediction and modeling of RO membranes performance is crucial in desalination processes for proper process control and operation. Existing models do not consider all process parameters, leading to less understanding of the parameter's importance. In this study, 5 non-ensemble and 7 ensemble machine learning models were employed to predict the performance of RO membrane. Data from a modern RO desalination plant in the UAE was utilized for the models' building. Thirteen input parameters, including operational parameters, water characteristic parameters, and time-dependent parameters, were used to predict salt rejection. The results suggested that ensemble techniques are more capable of predicting the performance of RO membranes. Among ensemble methods, the XGBoost model was found to outperform other models. Recursive feature elimination was integrated with Shapley additive explanation analysis to gain insights into the most influential predictors and confirm the model's ability to comprehend the RO membrane mechanism. The findings highlighted that five parameters are critical for predicting RO membrane performance and could be prioritized for future monitoring to provide timely membrane performance warnings: the membrane's age, feed water temperature, pressure, feed water flow, and chloride concentration. It also indicated that maintaining lower temperatures, increasing feed water pressure, and increasing feed flow can improve process efficiency. The optimal XGBoost model was found to have an outstanding predictive performance with a high R2 (94.75) and a low RMSE (0.181). Ultimately, the framework proposed by this study can serve as a tool to simplify and understand complex membrane processes.