Operational Improvements of Ultrafiltration Treatment of RO Feedwater Driven By Neural Network Models of UF Fouling and Backwash

Ultrafiltration treatment of reverse osmosis (RO) feedwater has been widely practiced in both inland and seawater desalination. However, UF requires periodic cleaning via backwash and eventually chemical in-place (CIP). Effective control and UF operation to enhance process performance and decrease frequency of CIP requires accurate system self-learning (for reasonable short-term forecasting and causal analysis) of the progression of fouling and backwash efficacy for the source water being treated. Admittedly, membrane fouling is a complex phenomenon governed by UF system design, operating conditions, water quality and temperature which can be temporally variable. In order to quantify UF operation and fouling progression under field conditions that are typically vary temporally, various fouling indicators have been proposed and evaluated with respect to quantification of UF backwash efficacy, UF resistance and post-backwash resistance. In the present work, UF system behavior is modeled via machine learning approaches, based on extensive data generated from field studies of UF-RO seawater desalination. UF operational models were developed based on artificial neural networks (ANN) and Alopex-based evolutionary algorithm (AEA) to predict fouling behavior of ultrafiltration (based on inside-out multichannel elements). The models were trained with data collected from a field study (at Port Hueneme, California) with an integrated UF-RO system (capacity of up to ~18,000 gallons/day) that utilized 8-inch RO elements and three inside-out polyethersulfone (PES) multi-bore hollow fiber UF membranes (0.02 μm pore size each module having an area of 50 m²). The neural network was trained considering various input variables including, for example, pressure, flow rate, temperature, and feed turbidity and chlorophyll concentration. Initially, an ANN model with AEA and Bayesian theory was developed to predict UF resistance, including both filtration and backwash modes for each operational cycle, with the outcome being the filtration and backwash performance for different water quality conditions. The analysis also included the impact of coagulant dose to the UF modules. An additional ANN model that includes AEA was also developed to predict the UF post-backwash resistance based on training of data collected during a storm event and where the coagulant dose was adjusted via a unique coagulant dose controller developed previously at UCLA. The ANN model with AEA demonstrated excellent performance (RMSE~0.0099, R²~0.947) in being able to properly account for backwash efficiency and prediction of the UF post-backwash resistance. The above modeling approach, as illustrated in the present work, provides a foundation for implementing improved self-adaptive operation of UF operation under variable field conditions.