Title: Neural network approach for modeling the performance of reverse osmosis membrane desalting
Journal: Journal of Membrane Science
Authors: Dan Liboteana, Jaume Giralta, Francesc Giralta, Robert Rallob, Tom Wolfec and Yoram Cohend
Corresponding author: Yoram Cohen
aFenomens de Transport, Departament d’Enginyeria Quimica, Universitat Rovira i Virgili, Av. Pa?sos Catalans 26, 43007 Tarragona, Catalunya, Spain
bFenomens de Transport, Departament d’Enginyeria Informatica i Matematiques, Universitat Rovira i Virgili, Av. Pa?sos Catalans 26, 43007 Tarragona, Catalunya, Spain
cToray Membrane USA, 12233 Thatcher Court, Poway, CA 92064, United States
dWater Technology Research Center, Chemical and Biomolecular Engineering Department, 5531 Boelter Hall, University of California, Los Angeles, CA 90095-1592, United States
The original and creativity of paper: Researchers simulated ANN models, which simulated with back propagation and support vector machines algorithm, using data obtained from RO pilot plant. The model described temporal variations in permeate flux and salt passage with a unique capability for useful short-term forecasting of process performance degradation.
This paper used experimental data from RO pilot plant for building the ANN models. The 1 MGD (million gallons per day) RO brackish water desalination plant was of a 2:1 array configuration located at Port Hueneme, CA, operated at 75% recovery. The monitored plant parameters for the feed stream included flow rate, conductivity, feed pressure, pH and temperature. Permeate and concentrate monitored parameters included flow rate, conductivity and pressure. These data were used as input variable for modeling.
ANN models were built for the standardized permeate flux and percent salt passage using back propagation (BP) and support vector machines (SVM) algorithm. The objective of this study is establishing the relationships between the selected inputs and target variables. In order to achieve the goal, an actual state-of-the-plant models (ASP) model and two types of forecasting models (sequential forecasting and matching forecast) for permeate flux and salt passage were first developed to assess the consistency of training and testing data sets, and the influence of process variables and algorithms in model building. Moreover, SOM was used to classify the complete operational data time-series using the five selected input process variables (feed flow rate, feed conductivity, feed pH, overall pressure drop, 2nd stage pressure drop) together with either the normalized permeate flux or salt passage targets.
The evaluations of standard time-series correlation (STSC) approach shows that model forecasting was limited to a period of about 30 min and required six and seven consecutive time-instants of previous permeate flux and salt passage data, respectively. These STSC models captured plant performance fluctuations over short periods but could not describe longer-term behavior. Therefore, the STSC modeling approach was deemed unsuitable for the development of practical plant diagnostics tools and early warning systems.
Analysis of various BP and SVR architectures and time-intervals demonstrated that it is feasible to model plant performance to a reasonable level of accuracy with respect to both permeate flux and salt passage, with a short-term memory interval of up to about 24 h. The present study focused on performance variability that was of short duration. Notwithstanding, the results of the present study suggest that there is merit in applying the present approach to plant operations that may involve longer time scales of performance degradation as is the case when membrane fouling and scaling occur. Current work is ongoing to extend and incorporate data-driven neural network-based models in a control strategy and early warning system of the deterioration of RO plant performance.
Application & further study: BP and SVR algorithms can be introduced to the RO lab-scale unit to evaluate and forecast system performance. As well as, this approach possible to forecast fouling and scaling which useful as control strategy and early warning system.
By Monruedee Moonkhum