AI-based prediction of total dissolved solids in reverse osmosis desalination for arid regions

Despite the increasing deployment of reverse osmosis (RO) desalination systems, predictive tools tailored to complex and variable operational conditions remain limited, particularly in arid regions. This study explores the use of artificial neural networks (ANNs) to predict total dissolved solids (TDS) in treated water across RO plants in the Ouargla region of Algeria. A dataset of 223 samples was collected over a 12-month period from eight desalination stations. Using this data, several ANN architectures were developed and evaluated. The optimal configuration featured a single hidden layer with 7 neurons. It achieved superior predictive accuracy compared to conventional models such as multiple linear regression and support vector regression. The model was trained using key operational parameters and validated for generalisation. It effectively captured the nonlinear interactions among variables, offering a reliable tool for performance optimisation and operational monitoring. Data augmentation techniques, including 1.5 % noise injection and a 1:0.5 synthetic-to-original ratio, were used to improve robustness under data-constrained conditions. This contributed to improved learning and model stability. Although the test-phase performance was modest, it reflected the best balance among all tested configurations and avoided severe overfitting observed in deeper architectures. These findings provide important guidelines for ANN design in water treatment contexts. Overall, the study contributes a validated, AI-driven modelling framework that supports efficient and resilient management of RO desalination systems. It demonstrates a useful case on the potential of machine learning to enhance water treatment operations in semi-arid environments.