(701e) Machine Learning-Based Model Predictive Control of a Closed-Circuit Reverse Osmosis System

Conventional reverse osmosis (RO) and membrane processes are widely employed for water desalination and purification, yet they suffer from high energy consumption, membrane fouling, mineral scaling, and limited water recovery [1-2]. These challenges lead to increased maintenance and operational costs, prompting the exploration of novel configurations. One promising alternative is the closed-circuit reverse osmosis (CCRO) process, where the production of permeate and disposal of brine are decoupled into distinct filtration and drain phases [3]. In the filtration phase, concentrate is recirculated to the membrane inlet, gradually increasing feed salinity until a preset threshold is reached, at which point the drain phase is triggered to flush the system with fresh feed water. Although this cyclic operation enhances energy efficiency and permeate quality, its dynamic and nonlinear behavior introduces significant control challenges.

Model Predictive Control (MPC) offers a substantial improvement over traditional PID controllers by forecasting future process behavior and optimizing control actions under multivariable constraints. However, implementing MPC in water treatment is limited, especially in niche industries where specialized expertise in model development, optimization, and control theory is scarce [4]. The accuracy of MPC is critically dependent on the underlying process model, and obtaining a first-principles-based, stable, high-fidelity differential-algebraic equation (DAE) model for CCRO is highly challenging.

Recent advances in machine learning, particularly long-short term memory (LSTM) neural networks, provide a viable alternative for capturing complex process dynamics [5-6]. This paper proposes integrating an LSTM-based model with MPC to develop a digital twin of the CCRO process. Trained on pilot-plant data, the LSTM model accurately predicts key variables such as concentrate conductivity and pump power consumption, thereby enabling predictive control and real-time optimization of the system. Comparison with an existing data-driven method is used to demonstrate the effectiveness of the machine learning model and its improved performance in MPC.

References

[1] Salinas-Rodríguez, S. G., Schippers, J. C., Amy, G. L., Kim, I. S., & Kennedy, M. D. (2021). Seawater reverse osmosis desalination: Assessment and pre-treatment of fouling and scaling. IWA Publishing.

[2] Zhu, A. (2012). Energy and cost optimization of reverse osmosis desalination. University of California, Los Angeles.

[3] Efraty, A. (2010). U.S. Patent No. 7,695,614. Washington, DC: U.S. Patent and Trademark Office.

[4] Rivas-Perez, R., Sotomayor-Moriano, J., Pérez-Zuñiga, G., & Soto-Angles, M. E. (2019). Real-time implementation of an expert model predictive controller in a pilot-scale reverse osmosis plant for brackish and seawater desalination. Applied Sciences, 9(14), 2932.

[5] Meng, J., Li, C., Tao, J., Li, Y., Tong, Y., Wang, Y., ... & Du, J. (2023). RNN-LSTM-based model predictive control for a corn-to-sugar process. Processes, 11(4), 1080.

[6] Han, Y., Ding, N., Geng, Z., Wang, Z., & Chu, C. (2020). An optimized long short-term memory network based fault diagnosis model for chemical processes. Journal of Process Control, 92, 161-168.

Acknowledgement

This material is based upon work supported by the National Alliance for Water Innovation (NAWI), funded by the U.S. Department of Energy, Office of Energy Efficiency and Renewable Energy (EERE), Industrial Efficiency & Decarbonization Office (IEDO) and was carried out at Oak Ridge National Laboratory under Contract No. DE-AC05-00OR22725 with UT-Battelle, LLC.