(216a) Rapid Prototyping of Reverse Osmosis Processes Using Data-Driven Models | AIChE

(216a) Rapid Prototyping of Reverse Osmosis Processes Using Data-Driven Models


Pascual, X. - Presenter, Universitat Rovira i Virgili
Gu, H. - Presenter, University of California, Los Angeles
Bartman, A. - Presenter, University of California, Los Angeles
Rahardianto, A. - Presenter, University of California, Los Angeles
Giralt, J. - Presenter, Universitat Rovira i Virgili
Rallo, R. - Presenter, Universitat Rovira Virgili
Giralt, F. - Presenter, Universitat Rovira i Virgili

Optimal control and optimization of reverse osmosis water desalination plants under conditions of changing water feed salinity, temperature and pH requires accurate RO process models that are specific to the plant under consideration. In this regard data-driven models of plant operation can be particularly useful for plant optimization and control, forecasting of plant performance, and assessment of operational deterioration (e.g., due to fouling and mineral scaling). In order to assess the adequacy of data-driven models of RO plant operation, the present work focused on the development and evaluation of such models using real-time data generated from a novel mini-mobile-modular (M3) RO pilot plant designed and built at UCLA. The experimental operating conditions covered a wide range of process parameters that included, for example, product water recovery, feed salinity, temperature, pH, and pressure and feed cross flow velocity. The fully automated M3 RO plant was programmed to autonomously and rapidly generate, via the M3 embedded computer controlled system, random and prescribed cyclical process conditions with the system allowed to both reach steady state for each reference condition, as well as operate under unsteady state conditions. The M3 RO plant was operated in a closed-loop configuration (i.e., RO brine and permeate returned to the feed tank during operation). In all cases steady state conditions were first established as a reference and perturbations were then invoked. Additional experiments were performed changing the salinity and the temperature of the feed stream to account for variability in the feed conditions. The collected data (recorded at high frequency) was then utilized to develop data-drive models based on Self-Organizing-Maps (SOM) and Support Vector Machines (SVM) to characterize the different operating regimes and to identify the transition paths between different plant states. Identification and quantification of the characteristic operating trajectories served as a basis for arriving at specific optimal operational criteria (e.g. energy savings and permeate productivity) while meeting with operating restrictions (e.g., avoidance of membrane fouling and mineral scaling). The present study demonstrated that data-driven RO plant models can be developed with a reasonable level of accuracy, and can be used in tandem with deterministic models to describe the complete domain of plant operability under both steady and unsteady state conditions.