A Multi-Model Ensemble for Advanced Prediction of Reverse Osmosis Performance in Full-Scale Zero-Liquid Discharge Systems.

The growing reliance on reverse osmosis (RO) in zero liquid discharge (ZLD) and seawater desalination has underscored membrane fouling as a critical challenge, requiring predictive tools for proactive management. This study proposes a novel multidimensional machine learning (ML) framework for forecasting RO performance in industrial ZLD systems. The framework includes data acquisition, feature engineering, ML modeling analysis, multidimensional evaluation, and integrated decision-making, which collectively enable accurate forecasting of fouling-related trends through the prediction of flux and salt rejection. Six ML models were assessed, and the convolutional long short-term memory (ConvLSTM) network exhibited superior performance for midterm (7 d, = 0.942) and short-term (1 d, = 0.960) predictions, capturing spatial and temporal dynamics. For long-term (30 d) forecasting, LSTM and ConvLSTM models achieved comparable performance, confirming suitability for extended prediction horizons. External validation across multiple industrial scenarios demonstrated the adaptability of the framework, enabling selection of optimal models for reliable predictions under diverse operational conditions. These findings demonstrated the capability of the framework to support proactive operational adjustments in response to fouling trends and enhance RO system stability. This study highlights the value of data-driven strategies in supporting operational decisions for industrial wastewater reuse and sustainable ZLD applications.