MINIMIZING ENERGY CONSUMPTION IN REVERSE OSMOSIS MEMBRANE DESALINATION USING OPTIMIZATION-BASED CONTROL

2010 American Control Conference Marriott Waterfront, Baltimore, MD, USA June 30-July 02, 2010 ThB14.6 MINIMIZING ENERGY CONSUMPTION IN REVERSE OSMOSIS MEMBRANE DESALINATION USING OPTIMIZATION-BASED CONTROL Alex R. Bartman, Aihua Zhu, Panagiotis D. Christofides and Yoram Cohen Abstract This work focuses on the design and implementation of an optimization-based control system on an experimental reverse osmosis (RO) membrane water desalination process in order to facilitate system operation at an energy optimal condition. A dynamic nonlinear lumped-parameter model for the RO process is derived using first principles and the model parameters are computed from experimental data to minimize the error between model predictions and actual system response. This model, along with several equations for reverse osmosis system energy analysis, are combined to form the basis for the design of a nonlinear optimization-based control system. The proposed control system is implemented on UCLA s experimental RO system and its energy optimization capabilities are evaluated. I. INTRODUCTION In recent years, the interest in the use of reverse osmosis desalination has increased due to the energy efficiency and versatility of the process [1]. Water shortages in various areas of the world have necessitated further development of the reverse osmosis desalination process in order to provide clean drinking water to the people in these regions. When operating a reverse osmosis desalination process, it is imperative that the system conditions be monitored and maintained at appropriate set points in order to produce the required amount of clean, potable water while preventing system damage. Even with advances in reverse osmosis technology, it is still a complex task to control a reverse osmosis desalination system in the presence of a variablequality feed source. With the rising costs of energy, it is also desired to find methods to reduce the energy consumption of these reverse osmosis desalination processes. This task, in conjunction with feed quality disturbances, requires a robust process control strategy. Traditionally, proportional-integral (PI) control algorithms are used to regulate permeate water flow rates and adjust the system pressure in order to achieve a desired rate of clean water production [2]. While these methods are able to maintain a constant permeate production rate, the PI control schemes fail to directly account for variability in feed water quality as well as energy usage by the reverse osmosis Alex R. Bartman, Aihua Zhu, and Yoram Cohen are with the Department of Chemical and Biomolecular Engineering, University of California, Los Angeles, CA 90095 USA. Panagiotis D. Christofides is with the Department of Chemical and Biomolecular Engineering and the Department of Electrical Engineering, University of California, Los Angeles, CA 90095 USA. Panagiotis D. Christofides is the corresponding author: Tel: +1(310)794-1015; Fax: +1(310)206-4107; e-mail: pdc@seas.ucla.edu. process. For certain operating conditions, mineral salt scaling or fouling may begin to occur on the membrane surface, leading to a decline of the permeate water flux through the membranes due to blockage. Using these traditional algorithms, the system is adjusted (increased feed flow rate, or higher system pressure) in order to increase the permeate flux back to the desired level. This process can speed up the development of mineral salt scaling or fouling, causing irreparable harm to the membranes and eventual system shutdown. Nonlinear control strategies have been developed to minimize the effects of varying feed water quality and also to account and correct for various faults that may present themselves during the operation of a reverse osmosis desalination process [3], [4]. Furthermore, control methods using model-predictive control (MPC) and Lyapunov-based control have also been evaluated using computer simulations [5], [6], [7], [8]. Reverse osmosis system analysis using linear models and data-based models using step-tests to create approximate linear models have also been demonstrated [9]. Other control methods have been evaluated in the context of RO system integration with renewable energy sources [10], [11]. In addition to the work using advanced nonlinear control algorithms in situations with high variability in feed water quality, extensive effort has been devoted to the issue of reverse osmosis system energy consumption. It has been shown that the cost of energy can approach 45% of the total permeate production cost due to the fact that the system operation can require (in the case of seawater) very high feed pressures (around 1000 psi) in order to achieve the required permeate production [12], [13], [14]. Several methodologies have been developed in order to decrease the energy required by a reverse osmosis desalination system; these include work in reducing membrane permeability leading to lower required transmembrane pressure [15], [16], optimization of RO module and system configuration [17], and also the use of energy recovery devices. In the area of energy optimization, it has been recently shown that the specific energy consumption, or SEC (energy cost per volume of permeate produced), is useful as a metric to quantify reverse osmosis desalination system energy usage [18]. Within the SEC framework, the issues of unit cost optimization with respect to water recovery, energy recovery system efficiency, feed flow rate optimization, and optimization of the transmembrane pressure have been studied [18], [19]. In this work, feedback control is integrated with energy 978-1-4244-7425-7/10/$26.00 2010 AACC 3629

optimization algorithms in order to maintain RO system operation at energy-optimal conditions. First, a reverse osmosis desalination system model is derived from mass and energy balances; then, the model parameters are computed based on experimental system data using a series of step tests. Next, the system model is used in conjunction with the system energy usage analysis equations in [18], [20] to design a nonlinear optimization based controller. This controller uses multiple system variables and a user defined permeate production rate to calculate the optimal control system set-points in order to minimize the specific energy consumption of the reverse osmosis desalination system. The nonlinear optimization-based control system is implemented on UCLA s experimental RO system and the data from experiments is shown to correspond closely to the theoretical minimum energy curve. It is also demonstrated that as the system reaches higher recoveries (in this work, near the theoretical optimum point), the concentration of salt in the product water (permeate) stream reaches levels above the drinking water standards (assumed in this work to be 500 parts per million). Several reasons for these results are discussed, and solutions to this problem are proposed. II. RO SYSTEM MODEL In order to utilize model-based control algorithms, an accurate system model must be derived. A simple RO process is used as the basis for the model derivation, as seen in Fig. 1; this process captures the main characteristics of UCLA s experimental RO system. In this process, the feed water enters at atmospheric pressure and is pressurized by the high pressure pump. The pressurized feed stream is fed to the reverse osmosis membrane module, and is split into a clean water (permeate) stream, and the more concentrated brine (or retentate) stream. After the retentate stream exits the reverse osmosis module, it passes through an actuated retentate valve that can be used to adjust the brine stream flow rate and the system pressure. For the model RO process in Fig. 1, mass and energy balances are used to derive a differential equation for the evolution of the retentate stream velocity and an algebraic relation for the system pressure. In the derivation, it is assumed that the water is incompressible, all components are operated on the same plane (potential energy terms due to gravity are neglected), and the density of the water is assumed to be constant. Derivation using the mass and energy balances leads to the following differential equation: dv r dt = A 2 p A m K m V (v f v r ) + A p ρv π 1 A p e vr vr 2 2 V where v r is the retentate stream velocity, A p is the pipe cross-sectional area, A m is the active membrane surface area, K m is the overall mass transfer coefficient, V is the system volume, v f is the feed stream velocity, ρ is the fluid density, π is the osmotic pressure difference across the surfaces of the membrane, and e vr is the retentate valve resistance. The equation for the relationship between osmotic pressure and (1) concentration was developed in [18] and can be represented as: ln( 1 1 Y π = f os C ) feed (2) Y where f os is an empirically obtained constant (f os = 78.7) [18], C feed is the total dissolved solids concentration of the feed solution, and Y is the overall system recovery. This value for osmotic pressure difference is also important in the additional algebraic equation for system pressure (P sys ); also derived from the mass and energy balances: P sys = ρa p A m K m (v f v r ) + π (3) Using an overall steady-state mass balance, the permeate stream velocity (v p ) can be determined by: v p = A mk m ρa p (P sys π) (4) As mentioned above, the metric used to determine the energy usage of the reverse osmosis desalination system is the specific energy consumption, or SEC. The SEC is defined as [18]: SEC = P (5) Y where P is the transmembrane pressure (pressure difference between the feed side and permeate side of the membrane), and Y is the recovery. The recovery is defined as the ratio of permeate produced to the amount of water fed to the system (or Y = Qp Q f ). The SEC can also be normalized with respect to the osmotic pressure of the feed (π 0 ): SEC norm = SEC π 0 (6) In this model formulation for the evaluation of system energy usage, it is assumed that the rejection of the membranes is equal to unity. It is also important to model the correlation between valve resistance value and the valve setting as controlled by the experimental system of Fig. 3. The correlation used in this work is the same as the correlation computed in [4]; the valve position is related to the valve resistance by three logarithmic relations, each used in a different range of valve resistance values. III. OPTIMIZATION-BASED CONTROL FOR SPECIFIC ENERGY CONSUMPTION MINIMIZATION After the model has been fit to the experimental step-test data and the model parameters have been determined as in [4], the energy-optimal controller is designed in order to facilitate system operation at the minimum specific energy consumption. Following [4], the experimental system uses LabVIEW software to record sensor data in real time. This LabVIEW program is interfaced with a MATLAB code that performs the energy optimization, then sends the controller/actuator set-points back to the LabVIEW software to be implemented on the system. As seen in Fig. 2, the 3630

Fig. 1. Simplified reverse osmosis (RO) system used in model development RO system user interface receives measurements of feed conductivity, feed flow rate, feed pressure, retentate flow rate, and retentate concentration from the process sensors. Using several of the values read from the sensors, the MATLAB code receives the data from the LabVIEW user interface and carries out the constrained optimization algorithm (see Eq. 13 below). When the optimal values for feed flow rate setpoint and valve position are determined, these values are passed back to the LabVIEW user interface and implemented on the system. It is important to note that as a minimum permeate production rate is commonly a requirement for any RO desalination operation, the optimization algorithm is constrained to finding optimal control values that satisfy the user-defined permeate production rate set-point. As it can also be seen from Fig. 2, the optimal feed flow rate (one of the pieces of data transmitted from the MATLAB optimization code back to the LabVIEW interface) is used as a set-point for a PI controller which regulates the speed of the pump using a variable frequency drive (VFD). The other value transmitted back to the user interface is the retentate valve setting; this value is directly implemented on the actuator. TABLE I PROCESS MODEL PARAMETERS BASED ON EXPERIMENTAL SYSTEM DATA. ρ = 1007 kg/m 3 V = 0.6 m 3 A p = 0.000127 m 2 A m = 15.6 m 2 K m = 6.4 10 9 s/m The objective of the optimization algorithm is to determine the values of feed flow rate (v f ) and retentate valve resistance (e vr ) such that the SEC at the operating condition is minimized; this can be represented as: min SEC (7) v f,e vr The optimization process follows several steps: 1) Raw sensor data retrieval from LabVIEW interface 2) Conversion of valve position to valve resistance value 3) Setting initial guesses for feed flow rate and valve position set-point 4) Optimization of control values at the given permeate flow rate using steady-state model (subject to constraints, see Eqs. 8-12) 5) Integrating model ahead one timestep using optimal values from previous step 6) Conversion of valve resistance value to valve position 7) Data transmission of controller/actuator set-points to LabVIEW interface In the raw data retrieval step, a UDP port is opened to allow for data transmission between the programs. The LabVIEW interface sends sensor measurements of the feed conductivity, feed flow rate, valve position (converted to valve resistance value as described in [4]), and permeate flow rate set-point to the MATLAB code approximately every 10 seconds. This timestep is dependent on the time taken to conduct the optimization; after the first step, each iteration takes between 1-5 seconds. This timestep can be easily tailored to a different RO system where optimization may take longer due to larger disturbances in the feed or faster changes in system set-points. The current values for feed flow rate and valve resistance (converted from current valve position) are used as the initial values for the first iteration of the optimization algorithm; after the first iteration, the previous optimal values are used in order to provide faster convergence. After the initial values are set, they are passed to the optimization function. The optimization function uses a steady-state version of the system model (due to the fast dynamics of the RO system) to determine the resulting system states and then calculates the resulting specific energy consumption (SEC) at each set of control values (through sequential quadratic programming). During the optimization, the control values (v f,e vr ) and resulting calculated system variables are subject to various constraints. The first constraint dictates that a constant permeate production rate is ensured: v p = v set p (8) where vp set is the permeate velocity set-point. For future work, it will also be possible to add a constraint on permeate conductivity (quality) so it is ensured that the system produces drinkable water. The next constraints ensure that the actuator set-points are positive (negative values have no physical meaning in the experimental system, and the output of negative values may damage the actuators). For 3631

Fig. 2. System control diagram detailing data flow between system sensors, system controllers, LabVIEW RO system user interface, and MATLAB optimization code the feed flow rate constraint, it is assumed that the feed flow rate is greater than or equal to the permeate flow rate for any reasonable operating condition (no back-flow into the modules through the retentate stream). These constraints dictate that the feed flow rate must be positive, as well as the valve resistance (and therefore, the valve position). v f 0 (9) e vr 0 (10) It is also necessary to constrain the SEC values to be positive in order to achieve the correct optimal control values. This constraint is represented as: SEC 0 (11) Additionally, it is required that the system pressure is greater than the osmotic pressure at the exit of the RO module. If this condition is not satisfied, part of the membrane surface area near the exit region of the module is not utilized to produce permeate water; due to the low applied pressure, the flow across the membrane in this region will actually be reversed (permeate water flowing back into the feed stream). Operation in this region (transmembrane pressure at the exit region below the osmotic pressure difference) is undesirable, and the process is constrained to operate at or above this limit. This constraint is also called the thermodynamic restriction as presented in [18]. P sys π 0 (1 Y ) (12) In summary, the optimization problem that yields the optimal control values for the feed flow rate and retentate valve resistance can be formulated as follows: min SEC v f,e vr v p = vp set v f 0 e vr 0 SEC 0 P sys π 0 (1 Y ) (13) As stated above, the optimization algorithm yields a setpoint feed flow rate to be used by the feed pump PI controller. The PI controller uses measurements of the feed flow rate to adjust the variable frequency drive (VFD) settings in order to achieve the desired feed flow rate set-point. The controller takes the form: V FD set = K f (Q sp f Q f) + K f τ f t 0 (Q sp f Q f)dt (14) where V FD set is the variable frequency drive setting. In this work, K f = 0.05 and τ f = 0.025. Finally, with the VFD PI controller operating, the RO system user interface applies the actuated valve position retrieved from the optimization code to the actuated retentate valve on the experimental system. IV. EXPERIMENTAL SYSTEM The experimental reverse osmosis water desalination system constructed at UCLA s Water Technology Research (WaTeR) Center was used for conducting the control experiments. This experimental system is comprised of a feed tank, two low-pressure feed pumps in parallel which provide enough pressure to pass the feed water through a series of cartridge filters while also providing sufficient pressure for operation of the high-pressure pumps, two high-pressure pumps in parallel (each capable of delivering approximately 4.3 gallons per minute at 1000 psi), and a bank of 18 pressure vessels containing Filmtec spiral-wound RO membranes. The 3632

high-pressure pumps are outfitted with variable frequency (or variable speed) drives which enable the control system to adjust the feed flow rate by using a 0-10V output signal. The bank of 18 membranes are arranged into 3 sets of 6 membranes in series; and for the control experiments presented below, only one bank of 6 membrane units was used (in the model, it is assumed that the 6 membranes in series can be represented by one RO module). After the membrane banks, an actuated valve is present to control the cross-flow velocity (v r ) in the membrane units, while also influencing system pressure. This valve is used as an actuator for the control system utilizing the control algorithms presented in section 3. The resulting permeate and retentate streams are currently fed back to the tank in an overall recycle mode, but for field operation the system can be operated in a one-pass fashion. The experimental system also has an extensive sensor and data acquisition network; flow rates and stream conductivities are available in real-time for the feed stream, retentate stream and permeate stream. The pressures before each high pressure pump, as well as the pressures before and after the membrane units (feed pressure and retentate pressure) are also measured. The system also includes sensors for measuring feed ph, permeate ph, in-tank turbidity, and feed turbidity after filtration (in real-time). A centralized data acquisition system takes all of the sensor outputs (0-5V, 0-10V, 4-20mA) and converts them to process variable values on the local (and web-accessible) user interface where the control system is implemented. The data is logged on a local computer as well as on a network database where the data can be accessed via the internet, while the control portion of the web-based user interface is only available to persons with proper authorization. The data acquisition and control system uses National Instruments software and hardware to collect the data at a sampling rate of 10 Hz and perform the necessary control calculations needed for the computation of the control action to be implemented by the control actuators. A labeled photograph of the system and its components can be seen in Fig. 3. V. RESULTS AND DISCUSSION In the experiments conducted in this work, the system was turned on and the PI controller for feed flow rate was activated. For these experiments, the permeate flow (Q p ) was set at a value of 1 gpm. The feed solution contained 3500 ppm of NaCl dissolved in deionized water. After the system reached a steady state, the nonlinear optimization MATLAB code was activated to begin transmitting the optimal setpoint values to the RO system interface. After the set-point values were received by the RO user interface, the set-points were implemented on the actuated retentate valve and the PI controller on the feed pumps. The sensor data taken from the experimental system was averaged (before transmission to the MATLAB code) using a 19 point moving average to remove the majority of the sensor noise. To obtain the other experimental sub-optimal data points, the system was manually adjusted (the optimization only provides the opti- Fig. 3. UCLA experimental RO membrane water desalination system: (1) Feed tank, (2) Low-pressure pumps and prefiltration, (3) High-pressure positive displacement pumps, (4) Variable frequency drives (VFDs), (5) Pressure vessels containing spiral-wound membrane units (3 sets of 6 membranes in series), (6) National Instruments data acquisition hardware and various sensors. mal set-point values for the actuator/controller) to achieve a range of feed pressures and feed flow rates while maintaining the desired constant permeate flow rate. When the system is operating at a fixed permeate flow rate, there exists only one degree of freedom with respect to the operating point. If the feed flow rate is changed, the pressure must take on a specific value to ensure that the permeate flow rate remains constant. The converse is true; if the system pressure is changed, the feed flow rate must take on a specific value in order to maintain the desired permeate flow rate. Because of this, for each normalized permeate flow rate, a single curve exists to describe the specific energy consumption at various recovery values [18]. The set of experiments presented in Fig. 4 demonstrates the experimental system s performance compared to the theoretical system performance. In the results, the SEC is normalized to SEC norm as discussed in Section II. From Fig. 4, it can be seen that the experimental system operating points are very close to the theoretically predicted operating points (in terms of specific energy consumption), for the ideal case of 100% salt rejection by the membranes. The minimum point in the experimental data also occurs very close to the theoretically-computed minimum. It can be seen that no points are tested at higher recoveries in order to demonstrate that this point is, in fact, the minimum of the experimental system SEC; this is due to the fact that it is not possible to achieve these recoveries with the current system setup due to physical limitations of the system components, while maintaining the desired permeate flow rate set-point. Another issue with the theoretical and experimental minimum points is that the resulting permeate salt concentration is above the maximum allowable permeate salt concentration. As mentioned previously in this work, the water produced must maintain a salt concentration below 500 ppm to be consistent with drinking water standards. As seen in Fig. 5, the permeate salt concentration increases with 3633

10 10 9 9 8 8 SEC norm 7 6 SEC norm 7 6 5 5 4 4 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Fractional Water Recovery Fig. 4. RO system specific energy consumption with respect to fractional water recovery at a fixed permeate flow rate; the dashed line represents the theoretical operating curve at a permeate flow rate of 1 gpm (assuming 100% salt rejection by the membranes), the diamonds represent experimental system data, and the vertical dotted line represents the permeate salt concentration limit of 500 ppm. Permeate Salt Concentration (ppm) 550 500 450 400 350 300 0.3 0.4 0.5 0.6 0.7 0.8 Fractional Water Recovery Fig. 5. Permeate salt concentration with respect to fractional water recovery at a fixed permeate flow rate; the diamonds denote experimental system data and the dotted line represents the permeate salt concentration limit of 500 ppm. system recovery (while maintaining a constant permeate flow rate) and passes the threshold value of 500 ppm at about 74% recovery; slightly below the optimum value of around 80%. This issue arises because the salt rejection (fraction of salt retained on the feed side of the membrane channel) of the membranes is not constant, as assumed in the optimization problem in Eq. 13. It has been found that the rejection of the experimental system decreases at increasing recovery values (while the constant permeate flow is maintained). Due to this fact, the salt concentration in the permeate stream increases with increasing recovery, which is observed in the data in Fig. 5. Since the rejection may be a complex function of the feed flow rate and of the transmembrane pressure, it is very difficult to include this expression in the model as a constraint since its explicit functional form 3 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fractional water recovery Fig. 6. RO system specific energy consumption with respect to fractional water recovery at a fixed permeate flow rate; the dashed line represents the theoretical operating curve at a permeate flow rate of 1 gpm assuming 100% salt rejection by the membranes, the solid line represents the theoretical operating curve at a permeate flow rate of 1 gpm accounting for membrane salt rejection, the diamonds represent experimental system data, and the vertical dotted line represents the permeate salt concentration limit of 500 ppm. is not available. The rejection is also dependent on (but not limited to) membrane structure, membrane composition, temperature, and the type of ions present in solution. Through these considerations, it can be seen that the rejection will be a complex function of the feed flow rate and of the transmembrane pressure that is unique to each reverse osmosis desalination system. If the rejection can be explicitly described in terms of feed flow rate and transmembrane pressure, then this expression can be used to theoretically determine the permeate concentration at any operating point. An explicit expression for rejection could also be used in the controller formulation as a constraint; in this way, the optimization code could determine the lowest SEC operating point that satisfies the permeate salt concentration standard. In the present work, the membrane rejections for the different water recoveries corresponding to the operating points of Fig. 4 were determined experimentally and used to recompute the theoretical operating curve at a permeate flow rate of 1 gpm. The resulting operating curve, accounting for membrane salt rejection, is shown in Fig. 6 as the solid line, and is very close to the experimentally computed operating points. It is noted that the deviation between the two curves increases at higher recoveries, owing to decreased membrane rejection. In the most likely situation (as in the current work), the expression for salt rejection dependence on process parameters is not known; therefore, the controller can not use the rejection (and subsequently, the permeate concentration) as an explicit constraint in the model-based optimization. Under these circumstances, additional controller operations must be defined in order to step back the system operating conditions from the recovery that gives the optimal SEC to a lower recovery value that provides permeate quality meeting 3634

the drinking water standards. It is proposed that the system can institute a procedure where, first, the variable frequency drive speed is increased (this will increase permeate flow rate, ensuring that the total permeate production stays at or above the required value); then, the system will open the retentate valve until the permeate flow rate drops back to the set-point value. In this way, the recovery should be decreased, while the salt rejection should increase, leading to a lower concentration of salt in the permeate stream. A feedback control loop measuring the permeate salt concentration can be used to determine when the permeate quality has reached the desired level; at this point, the stepping back procedure can be stopped. VI. CONCLUSIONS In this work, an experimental optimization-based control strategy was developed and experimentally implemented on a reverse osmosis (RO) membrane desalination system. First, a dynamic non-linear model of the system was derived using mass and energy balances. Next, this model was combined with a model for system specific energy consumption. After the model parameters were fit using experimental system step-test data, the controller was implemented on the system. During the system operation, sensor data and user defined set-points were sent from the RO system LabVIEW user interface to the MATLAB optimization code. In this code, the energy-optimal set-points for retentate valve position and feed flow rate were determined for the specified permeate flow rate. These set-points were then transmitted back to the RO system interface; the optimal feed flow rate value was then implemented as a set-point in the PI controller on the variable frequency drives, and the optimal retentate valve position was applied to the actuated retentate valve. The controller was shown to achieve energy usage values very close to the theoretically predicted values. It was also found that the salt rejection of the RO system decreased with increasing recovery (at a constant permeate flow rate); this issue can impose an additional constraint on the system operation due to high salinity in the permeate stream. The future research plans in this area include comparison of the energy optimal model-based controller to traditional reverse osmosis system controls in the presence of various disturbances; also, it is desired to design and implement methods to ensure permeate quality standards on the experimental reverse osmosis desalination system. [5] A. Abbas, Model predictive control of a reverse osmosis desalination unit, Desalination, vol. 194, pp. 268 280, 2006. [6] C. W. McFall, A. R. Bartman, P. D. Christofides, and Y. Cohen, Control and monitoring of a high recovery reverse osmosis desalination process, Ind. Eng. Chem. Res., vol. 47, pp. 6698 6710, 2008. [7] A. R. Bartman, C. W. McFall, P. D. Christofides, and Y. Cohen, Model predictive control of feed flow reversal in a reverse osmosis desalination process, J. Proc. Cont., vol. 19, pp. 433 442, 2008. [8] A. Gambier and E. Badreddin, Application of hybrid modeling and control techniques to desalination plants, Desalination, vol. 152, pp. 175 184, 2002. [9] I. M. Alatiqi, A. H. Ghabris, and S. Ebrahim, System identification and control of reverse osmosis desalination, Desalination, vol. 75, pp. 119 140, 1989. [10] D. Herold and A. Neskakis, A small PV-driven reverse osmosis desalination plant on the island of gran canaria, Desalination, vol. 137, pp. 285 292, 2001. [11] C. C. K. Liu, J. Park, R. Migita, and G. Qin, Experiments of a prototype wind-driven reverse osmosis desalination system with feedback control, Desalination, vol. 150, pp. 277 287, 2002. [12] T. Manth, M. Gabor, and E. Oklejas, Minimizing RO energy consumption under variable conditions of operation, Desalination, vol. 157, pp. 9 21, 2003. [13] M. Busch and W. E. Mickols, Reducing energy consumption in seawater desalination, Desalination, vol. 165, pp. 299 312, 2004. [14] M. Wilf and C. Bartels, Optimization of seawater RO systems design, Desalination, vol. 173, pp. 1 12, 2005. [15] M. Wilf, Design consequences of recent improvements in membrane performance, Desalination, vol. 113, pp. 157 163, 1997. [16] A. Zhu, P. D. Christofides, and Y. Cohen, On RO membrane and energy costs and associated incentives for future enhancements of membrane permeability, J. Membrane Sci., vol. 344, pp. 1 5, 2009. [17] W. G. J. van der Meer and J. C. van Dijk, Theoretical optimization of spiral-wound and capillary nanofiltration modules, Desalination, vol. 113, pp. 129 146, 1997. [18] A. Zhu, P. D. Christofides, and Y. Cohen, Effect of thermodynamic restriction on energy cost optimization of RO membrane water desalination, Ind. Eng. Chem. Res., vol. 48, pp. 6010 6021, 2009. [19] A. Zhu, P. D. Christofides, and Y. Cohen, Minimization of energy consumption for a two-pass membrane desalination: effect of energy recovery, membrane rejection and retentate recycling, J. Membrane Sci., vol. 339, pp. 126 137, 2009. [20] A. Zhu, P. D. Christofides, and Y. Cohen, Energy consumption optimization of reverse osmosis membrane water desalination subject to feed salinity fluctuation, Ind. Eng. Chem. Res., vol. 48, pp. 9581-9589, 2009. REFERENCES [1] A. Rahardianto, J. Gao, C. J. Gabelich, M. D. Williams, and Y. Cohen, High recovery membrane desalting of low-salinity brackish water: Integration of accelerated precipitation softening with membrane RO, J. Membrane Sci., vol. 289, pp. 123 137, 2007. [2] I. Alatiqi, H. Ettourney, and H. El-Dessouky, Process control in water desalination, Desalination, vol. 126, pp. 15 32, 1999. [3] C. W. McFall, P. D. Christofides, Y. Cohen, and J. F. Davis, Faulttolerant control of a reverse osmosis desalination process, in Proceedings of 8th IFAC Symposium on Dynamics and Control of Process Systems, vol. 3, Cancun, Mexico, 2007, pp. 163 168. [4] A. R. Bartman, P. D. Christofides, and Y. Cohen, Non-linear modelbased control of an experimental reverse-osmosis water desalination system, Ind. Eng. Chem. Res., vol. 48, pp. 6126 6136, 2009. 3635