The increasing scarcity of freshwater resources underscores the strategic importance of seawater desalination. However, optimizing reverse osmosis (RO) systems remains challenging because of raw water variability, high energy consumption, and membrane degradation. This study investigates the use of artificial intelligence (AI) for predictive monitoring of the Cap Djinet desalination plant (Boumerdès, Algeria), based on real operational data. Six supervised learning algorithms, linear regression (LR), polynomial regression (PR), support vector regression (SVR), random forest (RF), extreme gradient boosting (XGBoost), and multilayer perceptron (MLP), were evaluated for predicting the physicochemical and chemical parameters of the produced water. The findings indicate that ensemble models, particularly XGBoost (R2 = 0.999; RMSE = 6.11; MAPE = 2.23%), outperform other methods, followed by RF and SVR. Although simple, the LR model demonstrated strong robustness (R2 = 0.999; RMSE = 4.90), making it suitable for daily operation. The analysis further indicates that the performance of complex models, such as the MLP, is strongly influenced by the limited sample size (MAPE = 65.45%), illustrating the sensitivity of deep learning approaches in small-data contexts. While this frames the applicability of the results within the scope of the available dataset, it remains representative of the operational conditions commonly encountered in desalination plants with restricted yet meaningful datasets. Overall, XGBoost, RF, SVR, and LR demonstrate significant potential for predictive monitoring and sustainable optimization of desalination processes.
Mezhoud et al. (Thu,) studied this question.