SUPPLEMENT TO CHAPTER 12 FLOWSHEET CONTROLLABILITY ANALYSIS 12S.0 OBJECTIVES This chapter supplement introduces quantitative measures for controllability assessment to be used when developing the base-case design and in the detailed design stage (Stages 2 and 3, Table 12.1) and highlights how their integration into the design process can help to generate improved flowsheets that satisfy control performance criteria. At this point, the process creation stage has been completed and several promising process flowsheets exist. As they are evaluated, the control objectives are considered as constraints, the latter including: • Adequate disturbance resiliency, that is, the ability to reject disturbances quickly enough to meet specifications • Insensitivity to model uncertainty, that is, the ability to control easily, and to provide adequate closed-loop performance, with relatively insensitivity to model inaccuracies. An approach is introduced to screen the potential designs as early as possible, to identify the most promising designs for rigorous testing in stage 4, in which plantwide controllability assessment is completed. As demonstrated in this chapter supplement, it is important to verify the approximate analysis using rigorous dynamic simulation. Detailed multimedia instruction on the use of ASPEN HYSYS for dynamic simulation is available as part of the multimedia support that may be downloaded from the Wiley web site associated with this book. UNISIM and CHEMCAD could be used also for dynamic simulation. It is assumed that the reader is familiar with the basic concepts of linear systems theory. This material is covered typically in an introductory course on process dynamics and control at the undergraduate level. The subjects in that course that are prerequisite to understanding the concepts in this chapter supplement are: 1. Basic linear matrix theory, linearization, complex numbers, and Laplace and Fourier transforms. Note that Section 12S.1 provides some of this background material. – 12S-1– 2. Pole and zero positions in the complex plane, and their impact on the time-domain response of linear systems. 3. Linear stability theory and the impact of feedback. 4. Tuning of single-input, single-output controllers (P, PI, and PID controllers). Note that Section 12S.4 provides instruction on model-based PI-controller tuning. Key concepts relating to linear process models are reviewed in the first section of the chapter. For deeper coverage, the reader is referred to the following undergraduate-level texts: Bequette, B.W., Process Dynamics: Modeling, Analysis, and Simulation, Prentice Hall, Englewood Cliffs, NJ, (2003). Luyben, W.L., Process Modeling, Simulation and Control for Chemical Engineers, 2nd ed., McGraw-Hill, New York (1990). Ogunnaike, B.A., and W.H. Ray, Process Dynamics, Modeling and Control, Oxford Univ. Press, New York (1994). Seborg, D.E., T.F. Edgar, and D.A. Mellichamp, Process Dynamics and Control, Wiley, New York (1989). Stephanopoulos, G., Chemical Process Control, Prentice-Hall, Englewood Cliffs, NJ (1984). This supplement to Chapter 12, 1. Explains how to generate linear process models in their standard forms. 2. Defines quantitative measures that are used to analyze the controllability and resiliency (C&R) of process flowsheets, and shows how to implement them using MATLAB. 3. Describes a method to carry out C&R analysis using the results of steady-state process simulations. 4. Shows how to use quantitative analysis with steady-state and dynamic relative gain arrays (RGA and DRGA) to reliably select control loop pairings and to use the IMC model-based approach to provide preliminary tuning of single-loop PI controllers. 5. Analyzes, in Section 12S.5, selected case studies in Chapter 12 to demonstrate the utility of the quantitative methods. For completeness, these analyses are verified with dynamic simulations using single-loop PI controllers. – 12S-2– After reading this chapter, the student should 1. Be able to compute the frequency-dependent process transfer functions { P , P d } using MATLAB, given a linear model in one of its standard forms. 2. Generate the C&R measures: relative-gain array (RGA), and disturbance cost (DC), given the matrices P {s} and P d {s}, describing the effects of the manipulated variables and disturbances on the process outputs, using MATLAB. 3. Select the appropriate pairings for a decentralized control system for a process using the static and dynamic RGAs and appropriate resiliency measures, and provide preliminary tuning using IMC-PI tuning rules. 4. Perform C&R analysis to select between alternative process configurations, given the results of process simulations using linearized models. 12S.1 GENERATION OF LINEAR MODELS IN STANDARD FORMS In this chapter supplement, several methods are described to assist the designer in rejecting designs that do not provide acceptable closed-loop performance, using models linearized about a steady state. These are generated by expressing the open-loop response of the process outputs, y{s}, in terms of the variations of the inputs, u{s}, and disturbances, d{s}: y{s} = P{s}u{s}+ P d {s}d {s} (12S.1) The procedure for deriving the linear state-space model and the input-output transfer function model in Eq. (12S.1) involves the following steps: Step 1. The nonlinear state and output equations are derived from the material and energy balances that model the process. These are expressed in the form: dx = f {x, u, d } dt y = g {x, u, d } – 12S-3– (12S. 2) where x is a vector of nx state variables, y is a vector of ny output (measured) variables, u is a vector of nu manipulated variables, d is a vector of nd disturbances, and f and g are vectors of nx and ny nonlinear functions, respectively. Step 2. The state and output equations are solved at a stationary (steady) state that is defined either in terms of the desired state variable values or those of the input variables: f {x* ,u* , d * } = 0 y = g {x* ,u* , d * } (12S.3) where the stationary point is at x = x* , u = u* and d = d * . The solution of Eq. (12S.3) requires that the degrees of freedom of the process be resolved through the specification of nu+ nd values. Step 3. The equations are linearized in the vicinity of the desired stationary point, by a Taylor series expansion of Eq. (12S. 2): dx ≅ f { x* , u* , d *} + A ( x − x* ) + BU ( u − u* ) + B D ( d − d * ) + h.o.t. dt y ≅ g { x* , u* , d *} + C ( x − x* ) + DU ( u − u* ) + D D ( d − d * ) + h.o.t. (12S.4) Note that the linear approximation is obtained by ignoring the higher order terms (h.o.t.) of the Taylor series expansion. The matrices A, BU , B D ,C , DU and D D are the Jacobian matrices of appropriate dimension evaluated at the stationary point, defined as follows: { } A = ai , j ≡ { ∂f i ∂x j } BU = bU ,i , j ≡ { } B D = bD ,i , j ≡ { } C = ci , j ≡ x* ,u* ,d * ∂f i ∂u j ∂f i ∂d j { x* ,u* ,d * } DU = dU ,i , j ≡ { x* ,u* ,d * ∂g i ∂x j } D D = d D ,i , j ≡ x* ,u* ,d * ∂g i ∂u j ∂g i ∂d j x* ,u* ,d * x* ,u* ,d * Step 4. The linearized equations are formulated in terms of perturbation variables that express the deviation from the stationary point (or steady state): x̂ = x − x* , ŷ = y − y* , û = u − u* and – 12S-4– d̂ = d − d * . Substituting the perturbation variables into Eqs. (12S.4) and ignoring higherorder terms: d x̂ ≅ A x̂ + BU û + B D d̂ dt ŷ ≅ C x̂ + DU û + D D d̂ (12S.5) Eqs. (12S.5) constitute the linear state-space representation of the system. Step 5. The linearized equations are transformed into the Laplace domain: ŷ{s} = P{s}û{s}+ P d {s}d̂ {s}, (12S.1) where P{s} = C (s ⋅ I − A)−1 BU + DU and P d {s} = C (s ⋅ I − A)−1 B D + D D are matrices of the appropriate dimension. Eq. (12S.1) constitutes the input-output transfer function representation of the linear system. As an example, the procedure for generating linear models in standard form is demonstrated for an exothermic reactor, whose complete analysis is presented in Case Study 12S.1 of Section 12S.5. Example 12S.1 Standard Linear Models for an Exothermic Reactor. A continuous-stirred-tank reactor for the production of propylene glycol is analyzed in Case Study 12S.1, in Section 12S.5 below. Approximate linear models for the reactor are generated using the five-step procedure as follows: Step 1. Define the State and Output Equations. The hydrolysis of propylene oxide (PO) to propylene glycol is an exothermic reaction catalyzed by H2SO4: CH2-O-CH-CH3 + H2O → CH2OH-CH-OH-CH3 When water is supplied in excess, the reaction is second order with respect to the propylene oxide concentration and zero order with respect to the water concentration. Its rate constant exhibits an Arrhenius dependence on temperature, with k0 = 3.294×1026 m3/(kmol-h) and E = 1.556×105 kJ/kmol. Furthermore, it is customary to dilute the PO feed with methanol (MeOH), while the H2SO4 catalyst enters the reactor with the feed. Operating conditions are sought for carrying out – 12S-5– this liquid-phase reaction in a 47-ft3 continuous-stirred-tank reactor (CSTR), with the liquid holdup at 85% of its total volume (1.135 m3). The liquid feeds are fed at 23.9 oC, with one consisting of 18.712 kmol/h of PO and 32.73 kmol/h of MeOH. The water feed rate is from 160 – 500 kmol/h (2.84 – 8.88 m3/h), selected to moderate the reactor temperature. To reduce the risk of vaporization, the reactor is operated at a pressure of 3 bar. Under these conditions, the transients for the PO concentration, CPO [kmol/m3], and temperature, T [oC], are determined by solving the following species and enthalpy balances: dC PO ℑ PO ,in C PO (q0 + q w ) 2 = − − k {T }C PO dt V V (12S.6) 1 dT 2 (− ∆H ) − (q0 + q w )(T − T0 ) = k {T }C PO V dt C P (12S.7) where, k {T } = k o e − E / R (T + 273.2 ) m3/(kmol-h), R = 8.314 kJ/kmol-K, the molar flow rate of PO in the feed, ℑPO,in = 18.712 kmol/h, V = 1.135 m3, ∆H = –9×104 kJ/kmol, the organic volumetric feed rate, q0= 2.556 m3/h, the water volumetric feed rate is qw, T0 = 23.9 oC, and cP = 3,558 kJ/ m3 oC. Implicit in the assumption of perfect level control is the pairing between the effluent volumetric flow rate, F, and the liquid level, L. This leaves the temperature, T, as the output, to be controlled by the water feed rate, qw, as the manipulated variable. The disturbances to the process are the volumetric organic feed rate, q0, and the feed temperature, T0. Thus, x = [C PO ,T ]T , y = [T ] , u = [qw ] and d = [ℑ PO ,in ,To ]T . Step 2. Solve at the Steady State. The state equations are solved at the steady state. The degrees of freedom are resolved by fixing all of the input variable values and solving the two equations for the two unknown state variables. C PO (q0 + q w ) 2 − k {T }C PO =0 V (12S.8) 1 2 (− ∆H ) − (q0 + qw )(T − T0 ) = 0 k {T }C PO V CP (12S.9) ℑ PO ,in V − Taking u* = qw* = 5.325 m3/h and d * = [ℑ PO ,in* ,To ]T = [18.712 , 23.9]T and solving Eqs. (12S.8) and (12S.9) gives x* = [0.06 , 82.4]T . Note that the fractional conversion of PO is X = 1 – – 12S-6– CPO/CPO,in = 1 – 0.06/2.374 = 0.975, where CPO,in = ℑPO,in/(q0 + qw). This solution is obtained analytically, graphically (see Case Study 12S.1), or using a numerical method (e.g., the NewtonRaphson method). Steps 3 and 4. Linearize in the Vicinity of the Steady State. The Jacobian matrices for the linearized approximation are: (q0 + qw* ) − − 2k {T* }C PO* 1 V A= 60 (− ∆H ) 2k {T }C * PO* CP ∂k {T* } 2 C PO* ∂T 2 (q + qw* ) + ∂k {T* }(− ∆H )C PO* − 0 ∂T V CP − C PO* 1 0 − 1 1 V V BU = (T − T ) , B D = 60 − * 0* 60 0 (q0 + q w* ) V V C = [0 1], DU = [0] , D D = [0 0] (12S.10) (12S.11) (12S.12) The division by 60 in each matrix is to express time in minutes instead of hours. Note also that all of the variables are expressed in physical units. Substituting numerical values into Eqs. (12S.10) and (12S.11): − 9.203 − 0.0396 , A= 0.8870 229.8 0 − 0.0009 0.0147 , BU = BD = 0.1157 − 0.8596 0 These matrices are scaled by assuming that all outputs and manipulated variables are nominally at 50% of their ranges, and the disturbance variable values are constrained to vary in the range ∆d = [±50% ±5 oC]T. Thus: 0 C − 9.203 − 55.43 A s = S −x 1 AS x = , S x = PO* T* 0.1644 0.8870 0 − 0.0782 BU ,s = S −x 1 BU S u = , S u = q w* − 0.0555 0 2.330 0.5ℑ PO ,in* 0 B D ,s = S −x 1 B D S d = ,S d = 0.0070 0 5 0 – 12S-7– (12S.13) These matrices relate the input (manipulated and disturbance) variables to the output (controlled) variables, with all of the variables scaled and in perturbation variable form. Step 5. Generate Transfer Functions. These are computed using the scaled matrices in Eq. (12S.13), for example ( P{s} = C s ⋅ I − A s ) −1 −1 BU , s + DU , s 55.43 − 0.0782 s + 9.203 = [0 1] − 0.1644 s − 0.8870 − 0.0555 Hence, P{s} = − 0.0555s − 0.524 2 s + 8.32 s + 0.949 = − 0.552(0.106 s + 1) (0.122s + 1)(8.64s + 1) (12S.14) Note that P{s} is a scalar transfer function, since it relates perturbations in the single manipulated variable, q̂w , to those in the single process output variable, Tˆ . Note that the process zero almost cancels the fast process pole, meaning that the response to the manipulated variable is effectively that of a first-order lag, with a time constant of approximately 9 min. A similar computation yields the transfer function matrix, P d {s} : −1 55.43 2.330 0 s + 9.203 P d {s} = [0 1] 0.0070 − 0.1644 s − 0.8870 0 0.410 = (0.122s + 1)(8.64s + 1) 0.068(0.109s + 1) (0.122s + 1)(8.64s + 1) (12S.15) The columns of P d {s} define the responses of Tˆ to ℑ̂ PO ,in and T̂o , respectively. Note that the temperature response to step changes in ℑ PO ,in is of second order, while its response to changes in the feed temperature is effectively of first order. The reader is referred to the Wiley website that accompanies this text for useful MATLAB functions and scripts for the generation of linear models in their standard forms. – 12S-8– 21.2 QUANTITATIVE MEASURES FOR CONTROLLABILITY AND RESILIENCY The quantitative assessment of the controllability and resiliency of chemical processes has generated considerable interest. The term resiliency was introduced by Morari (1983), who also pioneered qualitative measures for its assessment. Furthermore, Perkins (1989) presented an approach for the simultaneous design of processes and their control systems that addresses plantwide controllability directly. All of the C&R measures use the linear approximations, P{s} and P d {s} , which describe the effects of the control variables and disturbances, respectively, on the process outputs. A commonly used controllability measure is the relative-gain array (RGA – Bristol, 1966), which relies only on P{s} . The disturbance condition number (DCN; Skogestad and Morari, 1987) and the disturbance cost (DC; Lewin, 1996) are resiliency measures that require a disturbance model, P d {s} , in addition to P{s} . These C&R measures are especially useful in Stages 2 and 3 of the design process (see Table 12.1) because they do not assume a controller structure or a specific controller design and tuning. It is assumed that each input variable is nominally at the midpoint of its range and is expressed in perturbation variable form, and scaled by dividing by its nominal value. For example, if Fi is an inlet flow rate, nominally at 500 lbmol/hr, its operating range is 0 ≤ Fi ≤ 1000 lbmol/h, in perturbation variable form, −500 ≤ Fi ≤ 500, and in scaled form, −1 ≤ Fi ≤ 1. Thus, P{s} and P d {s} are scaled by multiplying the gains in each column by the nominal value of the appropriate input variable. As a result, all of the scaled inputs vary over the same range [−1, 1]. Note, however, that the RGA is scale independent, whereas the DC is input scale dependent. – 12S-9– Relative-gain Array (RGA) Steady-State RGA (Bristol, 1966) Figure 12S.1 shows the block diagram for a multiple-input, multiple-output (MIMO) process to be controlled by two single-loop controllers. Having closed one of the loops (y1 − u1), the controller in the second loop, which manipulates u2 based on the feedback of y2, must be tuned. A desirable feature of this controller is to have the effective process gain remain invariant, regardless of the action of the other control loop. - c1 u1 p11 y1 p21 p12 u2 p22 y2 Figure 12S.1 MIMO process with one control loop. When the controller c1 is put into manual operation, i.e., when it is turned off, the process gain as seen by controller c2 is y2 = p22 (12S.16) u2 c ,OL 1 where u2 and y2 are the deviations of the input and output from their nominal values in the steady state. On the other hand, when c1 is put into automatic operation, the process gain as seen by controller c2 is y2 c1 p21 = p22 − p12 1 + p11c1 u2 c ,CL 1 In general, for MIMO systems, a useful measure is the ratio process gain as seen by a given controller with all other loops open process gain as seen by a given controller with all other loops closed – 12S-10– (12S.17) When this ratio is close to unity, the given controller is relatively insensitive to interaction. Computing this ratio for the MIMO process in Figure 12S.1: y2 u2 y2 u2 c1 ,OL = c1 ,CL p22 (12S.18) c1 p22 − p12 p21 1 + p11c1 When the top loop is closed-loop stable, and when c1 has integral action, lim c1 1 = s → 0 1 + p11c1 p11 Therefore, the ratio at steady state is y2 u2 lim s→0 y 2 u2 Similarly, c1 ,OL c1 ,CL = s lim →0 p22 c1 p22 − p12 p21 1 + p11c1 = p11 p22 p11 p22 − p12 p21 y2 lim u1 c1 ,OL − p12 p21 = s → 0 y2 p11 p22 − p12 p21 u1 c ,CL (12S.19) (12S.20) 1 Thus, for a two-input, two-output process, the RGA is defined as y1 y1 u1 c2 ,OL u2 c2 ,OL y1 y1 u u 1 c2 ,CL 2 c2 ,CL p11 p22 − p12 p21 λ11 λ12 −1 = Λ = = ⋅ det ( P ) − p12 p21 p11 p22 y2 λ 21 λ 22 y2 u1 c1 ,OL u2 c1 ,OL y2 y2 u1 c ,CL u2 c ,CL 1 1 In general, the RGA can be computed using ( ) Λ = P ⊗ P −1 T where ⊗ denotes the element-by-element (Schur) product. – 12S-11– (12S.21) (12S.22) Theorem (2 × 2 Systems Only). If λ11 (= λ22) is positive, there exists a pair of single-input, single-output (SISO) controllers, c1 and c2, with integral action for the loops u1 − y1 and u2 − y2 such that the loops are stable by themselves and together. If λ11 is negative, there are no controllers that can guarantee stability by themselves and together. In other words, to guarantee closed-loop stability with either of the two SISO controllers in automatic or manual, the controllers should be paired such that the RGA elements corresponding to the pairings are positive. Negative RGA elements are an indication of the presence of destabilizing positive feedback due to unfavorable process interactions. Similarly, excessively large RGA elements are related to poorly conditioned processes; those in which the effective process gain may be orders of magnitude different, depending on the input direction. For systems of higher rank, a necessary condition for the stabilizability of a decentralized control system is the selection of pairings such that λij > 0, and hence the RGA provides a useful screening tool. The decentralized integral controllability (DIC) conditions (see Morari and Zafiriou, 1989, pp. 359-367) provide additional necessary conditions for the stability of higherorder systems, which depend only on the steady-state gain matrix, P{0}. Properties of the Steady-state RGA The following properties are especially noteworthy when working with the RGA: 1. ∑ λij = ∑ λij = i 1 (rows and columns sum to unity) j 2. If P is triangular (lower or upper), Λ = I 1 0 0 − 3 2 3 e.g. P = 0 3 5 ⇒ Λ = 0 1 0 0 0 1 0 0 1 In such systems, the process interaction is in one direction only, and therefore, precludes the possibility of the occurrence of destabilizing feedback. 3. For 2×2 systems only: If P has an odd number of positive elements, 0 ≤ λij ≤ 1 If P has an even number of positive elements, either λij < 0 or λij > 1 – 12S-12– Dynamic RGA (McAvoy, 1983) Considering the same MIMO process in Figure 12S.1, y2 is expressed in terms of u1and u2: y2 = p21u1 + p22u2 (12S.23) When c1 is in manual operation, u1 = 0 and y2 = p22 u2 c ,OL 1 as for the steady-state analysis. When c1 is in automatic operation and it is assumed that the first loop can be designed to give perfect control (i.e., the first loop's output is assumed to be held at its set point), p y1 = p11u1 + p12 u2 = 0 ⇒ u1 = − 12 u2 p11 Substituting for u1 in Eq. (12S.23), (12S.24) y2 p p = p22 − 12 21 (12S.25) u2 c ,CL p11 1 Hence, the dynamic RGA (DRGA) has precisely the same form as the steady-state array. Note that the dynamic RGA assumes perfect control, which may not be an appropriate assumption, especially at high frequencies. The computation of the DRGA requires care since it involves complex algebra. Because columns and rows sum to unity only at the steady state, the DRGA should be computed using: ( ) DRGAij {ω} = sign λ ij {0} ⋅ λ ij { jω} , (12S.26) with λij{jω} computed conveniently for 2 × 2 systems using Eqs. (12S.19) and (12S.20), or using Eq.(12S.22) in general. An accepted rule of thumb is to avoid pairings between variables with negative RGA elements and to select those with values close to unity, as illustrated in the following example. – 12S-13– Example 12S.2 LV Control of a Binary Distillation Column Figure 12S.2 shows the LV configuration for the two-point composition control of a binary distillation column discussed in Example 12.9. After assigning manipulated variables to regulate the vapor and liquid inventories, the boilup rate, V, and the reflux flow rate, L, remain available to control the distillate and bottoms product compositions, xD and xB, respectively. To assess the controllability and resiliency of this configuration, the disturbances are taken to be the feed composition, xF, and the flow rate, F. The column dynamics are approximated by a linear model in transfer function form (Sandelin et al., 1990): −0.045 − 0.5 s 0.048 − 0.5 s −0.001 e − s 0.004 e − s F e L xD 8.1s +1 e 11s +1 10 s +1 8.5 s +1 (12S.27) x = − 0.23 −1.5 s 0.55 − 0.5 s V + − 0.16 − s − 0.65 − s x e e F e e B 8.1s +1 10 s +1 9.2 s +1 5.5 s +1 To complete the process model definition, it is noted that the process input ranges are as follows: L = 60 ± 60 kmol/h, V = 72 ± 72 kmol /h, F = Fnom ± 20 kmol /h, xF = xF, nom ± 6 %. In Eq. (12S.27), the gain coefficients are in the appropriate units, and time is in minutes. Figure 12S.2 Control of a binary distillation column using the LV configuration. The qualitative guidelines presented in Chapter 12 are not sufficient to decide how to pair the two manipulated variables with the two outputs. Without analysis, it is not clear whether this pairing should be diagonal (i.e., {xD−L, xB−V } as shown in Figure 12S.2) or off-diagonal (i.e., {xD−V, xB−L }). However, using Eq. (12S.19), λ11 in the RGA is – 12S-14– = λ11 p11 p22 1.8 = p11 p22 − p12 p21 (12S.28) Using the property that the RGA rows and columns add to unity, 1.8 − 0.8 Λ= , − 0.8 1.8 and consequently, diagonal pairing is recommended, with the off-diagonal pairing resulting in stability problems, either when both of the controllers are on automatic or when one of the controllers is switched to manual operation. Although stable, significant interactions are anticipated when both loops are closed, because of the large RGA element. Figure 12S.3 Closed-loop response of the LV configuration for binary distillation to the worstcase disturbance, d = [20, 6]T, with decentralized PI control - Outputs: xD (solid line), xB (dashed line); Inputs: L (solid line), V (dashed line). To verify this, Figure 12S.3 shows the closed-loop response for the process, diagonally paired with IMC-tuned PI controllers (xD − L loop: Kc = -50, τI = 8 min; xB − V loop: Kc = 5, τI = 10 min). For the IMC-PI tuning rules, the reader is referred to Section 12S.4. The simulation is computed for the worst-case disturbance, d = [20, 6]T, identified using the disturbance cost analysis, to be discussed shortly. As expected, the response is stable but shows significant interactions, with the bottoms composition affected more significantly. The reader can try out this example under MATLAB, using the interactive C&R Tutorial CRGUI available on the Wiley website that accompanies this text. In the main menu, opt for the “Binary Column.” In some cases, the RGA elements vary significantly with the frequency, which may indicate bandwidth limitations on the diagonal dominance of the process. For this example, Figure 12S.4 – 12S-15– shows λ11 and λ12 as a function of the frequency. Although there is considerable variation at high frequencies, the diagonal dominance holds for the entire frequency range of interest. In this case, the RGA and DRGA give the same pairing recommendations. For some processes, however, the information furnished by the dynamic RGA can be crucial for the correct pairing selection. The following example provides one such case. Figure 12S.4 Dynamic RGA for the diagonal (solid line) and off-diagonal (dotted line) pairings for Example 12S.2. Example 12S.3 Importance of the Dynamic RGA Consider the process: 2.5 − 5s y1 (15 s +1)(2 s +1) e y = 1 2 3 s +1 5 u − 4 e − 2s 4 s +1 1 + 10 s +1 −1 − 2 s −4 e − 5 s u 2 5 s +1 e 20 s +1 3 e − 5 s d1 10 s +1 −2 e − 5 s d 2 10 s +1 (12S.29) Here the process inputs are limited to the ranges: u1 = 60 ± 60, u2 = 50 ± 50, d1 = d 1,nom ± 20, and d 2 = d 2,nom ± 5, and time is in minutes. For this system, λ11 = 2/3 in the steady-state RGA, suggesting that the variables be paired diagonally. In the dynamic RGA, however, the diagonal dominance deteriorates at moderate frequencies, as shown in Figure 12S.5. In fact, the process is off-diagonally dominant in the frequency range of interest. The open-loop time constants are on the order of 10 min, and hence, frequencies in the range 0.1 < ω < 1 rad/min are of particular interest. For this system, the offdiagonal pairing (i.e., y1 − u2 and y2 − u1) is preferred, contrary to the pairing suggested by the steady-state RGA. To verify the analysis in Figure 12S.5, the two pairings are simulated using – 12S-16– IMC-PI tuning rules (see Section 21.4). For the diagonal pairing, the controllers are tuned: y1 − u1 loop, Kc = 0.6, τI = 15 min; y2 − u2 loop, Kc = −0.37, τI = 20 min. In contrast, for the off-diagonal pairing, the controller tuning parameters are: y2 − u1 loop, Kc = 10, τI = 3 min; y1 − u2 loop: Kc = 2, τI = 4 min. Note that the controller gains for the diagonal pairing are an order of magnitude lower than for the off-diagonal pairing, a reflection of the bandwidth limitations imposed by the delays on the diagonal elements of the process transfer-function matrix. For a unit-step increase in the y1 setpoint, Figure 12S.6 shows that although the diagonal controller is bandwidth limited, the tuning for the off-diagonal configuration can be arbitrarily aggressive, only restricted by the actuator constraints. The reader can try out this example under MATLAB, using the interactive C&R Tutorial CRGUI available on the Wiley website that accompanies this text. In the main menu, opt for the “Mystery Process.” Figure 12S.5 Dynamic RGA for the diagonal (solid line) and off-diagonal (dotted line) pairings for Example 12S.3. Figure 12S.6 Closed-loop response for Example 12S.3 with PI control for a setpoint change in y1 using: (a,b) diagonal pairing; (c,d) off-diagonal pairing. Shown on the top row are outputs: y1 (solid line), y2 (dashed line), and on the bottom row, inputs: u1 (solid line), u2 (dashed line). – 12S-17– Example 12S.4 Control Configuration for a Utilities Subsystem (Example 12.7 Revisited). The analysis of the utilities subsystem in Figure 12S.8 is based on the steady-state material and energy balances: Fc = Fc1 + Fc 2 Tco = (12S.30) Fc1Tc1 + Fc 2 Tc 2 Fc1 + Fc 2 , (12S.31) where the controlled variables are Fc and Tco. Selecting Fc1 and Fc2 as the manipulated variables, the steady-state gain matrix is computed by partial differentiation of Eqs. (12S.30) and (12S.31): 1 ∆Fc ∆T = Fc 2 (Tc1 − Tc 2 ) co (F + F )2 c2 c1 1 ∆F − Fc1 (Tc1 − Tc 2 ) c1 ∆F (Fc1 + Fc 2 )2 c 2 (12S.32) The λ11 element of the RGA is computed using Eq. (12S.19): λ 11 = p11 p22 p11 p22 − p12 p21 = Fc1 Fc1 + Fc 2 =x (12S.33) where 0 ≤ x ≤ 1. Since the RGA rows and columns add to unity, the RGA matrix is [ Λ = 1 −x x 1 −x x ] (12S.34) The recommended pairings depend on the operating conditions of the coolant subsystem, with significant interactions normally occurring between the control loops. To avoid this, the manipulated variables are defined as ϕ = Fc1+Fc2 and µ = Fc1/(Fc1+Fc2), transforming Eqs. (12S.30) and (12S.31) to Fc = ϕ (12S.35) Tco = µ ⋅ Tc1 + (1 − µ ) Tc 2 (12S.36) After partial differentiation, 0 ∆ϕ ∆Fc 1 ∆T = 0 T − T ∆µ c1 c2 co – 12S-18– (12S.37) This is a decoupled system that requires diagonal pairings, because Λ = I . Figure 12S.7 An attractive control configuration for the utilities subsystem. These pairings, shown in Figure 12S.7, are intuitively correct in that the total flow rate is controlled by the sum of the two utility streams, and the coolant temperature is controlled by the fraction of the coolant flowing through the heating system. Note that the temperature and flow controllers manipulate the variables µ and ϕ, respectively, which are processed by a decoupler, D, to generate corrections to the two flow rates, Fc1 and Fc2, according to: Fc1 = ϕ ⋅ µ Fc 2 = ϕ (1 − µ ) (12S.38) Example 12S.5 Control Configuration for a Debottlenecked Distillation Column Often, process design modifications can lead to potential control problems, as demonstrated by McAvoy (1983) for a distillation column in which the reboiler capacity is doubled by the addition of an identical reboiler in parallel with the original one (i.e., the column is debottlenecked), as shown in Figure 12S.8. Figure 12S.8 Debottlenecked distillation column. – 12S-19– A MIMO control system must be configured for the retrofitted column. To compute the RGA, a linearized model, in the steady state, relates the changes in the designated outputs, T, L1, and L2, to those of the manipulated variables, Q1, Q2, and B: ∆T a11 a12 ∆L = a 1 21 a22 ∆L2 a31 a32 a13 ∆Q1 a23 ∆Q2 a33 ∆B (12S.39) Since B does not affect T directly, ∆T ∆B = 0 . Furthermore, by symmetry, a11 = a12 = ∆T ∆T = ∆Q1 ∆Q2 a 23 = a33 = ∆L1 ∆L2 = ∆B ∆B a 21 = a32 = ∆L1 ∆L2 = ∆Q1 ∆Q2 a 22 = a31 = ∆L1 ∆L2 = ∆Q2 ∆Q1 Hence, Eq. (12S.39)) becomes ∆T 1 ∆L= ∆L2 a11 0 ∆Q1 a11 a 21 βa21 a23 ∆Q2 βa21 a21 a23 ∆B (12S.40) where = β a22 effect of Q1 on L2 = <1 a21 effect of Q1 on L1 Using Eq. (12S.22), 0.5 T = 0.5 1−β −0.5β 1−β ( ) Λ= P ⊗ P −1 0 0.5 0.5 0.5 −0.5β 1−β 0.5 1−β (12S.41) Note that β < 1 and is close to unity. Assuming β = 0.95, the RGA becomes 0.5 0.5 Λ = 10 − 9.5 − 9.5 10 0 0.5 0.5 (12S.42) To ensure no loss of stability, pairings on negative RGA coefficients are avoided. Thus, only two possible pairings remain to be considered: [Q2−T, Q1−L1, B−L2] and [Q1−T, B−L1, Q2−L2] . Neither – 12S-20– alternative gives good performance since, in each case, one loop has a relative gain of 10, implying the need for severely detuned controllers. Clearly, the pairing selection for both of these controller configurations is limited by the available outputs and manipulated variables, and does not exploit the symmetry in the process design. This drawback can be avoided by selecting other manipulated and controlled variables. Here, it is desired to control the total hold up (ψ = ∆L1 + ∆L2), and the best manipulated variable to do this is intuitively the bottoms flow rate, B. Thus, the vector of manipulated variables is redefined as u = [φ Γ ∆B]T, where φ = ∆Q1 − ∆Q2 and Γ = ∆Q1 + ∆Q2, and the vector of controlled variables is redefined as y = [∆T Ω ψ]T , where Ω = ∆L1 − ∆L2. The linear model, expressed in terms of these new variables, becomes Ω (1 − β )a21 ∆T = 0 ψ 0 0 a11 (1 + β )a21 0 φ 0 Γ 2a23 ∆B (12S.43) Note that this is a lower-triangular matrix, with the corresponding RGA: 1 0 0 Λ = 0 1 0 0 0 1 This result suggests the following pairings: (1) Ω − φ (the imbalance in the holdups controlled by the imbalance in the heat duties of the two reboilers), (2) ∆T − Γ (the reboiler temperature controlled by the total heat duty), and (3) ψ − ∆B (the total holdup controlled by the bottoms flow rate). These control loops largely respond independently of each other (there is small one-way interaction between the second and third loops), and are referred to as decoupled. The RGA as a Measure of Process Sensitivity to Uncertainty Thus far, the RGA has been used to measure the process interactions and to aid in selecting the pairings for decentralized controller configurations. It is noteworthy that the magnitudes of the RGA elements are an indication of the degree of the process sensitivity to uncertainty. This is illustrated using a hypothetical process model, K (1 − ε ) 1 , P= K 1 – 12S-21– (12S.44) in which the p11 coefficient is subject to a fractional uncertainty, ε. This uncertainty can significantly affect the RGA, depending on the value of K, as shown in Figure 12S.9, where = λ11 K 2 (1 − ε ) [ K 2 (1 − ε ) − 1] , is displayed as a function of ε for two values of K. For K = 10, the process is strongly diagonally dominant and hardly affected by the uncertainty ( λ11 is close to unity). On the other hand, for K = 2, λ11 = 1.33 when ε = 0, indicating that the process has significant interactions. Furthermore, P becomes singular at ε = 0.75 and the recommended pairings are switched, implying that a multivariable control system is unreliable at this level of uncertainty. Figure 12S.9 Effect of uncertainty on the RGA for K = 2 and K = 10. In summary, processes with RGA coefficients close to unity are relatively insensitive to uncertainties in the process model. Conversely, processes with large RGA coefficients tend to exhibit a high degree of sensitivity to model uncertainties. Using the Disturbance Cost to Assess Resiliency to Disturbances The design of process controllers for open-loop stable systems is motivated principally by the need to impart disturbance resiliency properties to processing operations. In other words, it is – 12S-22– intended to maintain the outputs of multivariable processes at their set points despite external disturbances and uncertainties in the process model. The degree to which this requirement is satisfied is referred to as resiliency. Given the process model of Eq. (12S.1) and assuming perfect control, the action required to completely reject the disturbance, d, is u{s} = − P −1{s}d ′{s}, where d ′{s} = P d {s}d {s} (12S.45) By computing the norm of the actuator response, u , as a function of the disturbance direction, the relative cost of rejecting a particular disturbance, d, is computed as a function of its direction. One quantitative measure of the control effort to reject a given disturbance vector is the Euclidean norm: u{s} 2 = P −1{s}P d {s}d {s} , 2 (12S.46) it being noted that the infinity norm provides an alternative resiliency measure. Parseval's theorem provides the direct translation of the 2-norm, in the frequency domain, to the total control action in the time domain. This norm, u 2 , is the disturbance cost (DC; Lewin, 1996). Often, it is more helpful to compute DC values for each manipulated variable separately. Since u 2 is a frequency dependent measure, it can be displayed as a function of frequency and the direction of d{s}, to show the effect of two disturbances d1 and d2, where the disturbance direction is the angle of the disturbance vector with respect to the abscissa, that is, arg{d}. Contour maps of DC are displayed as a function of the disturbance direction and frequency. Since the DC is based on the assumption of perfect control, the results are independent of controller tuning or sophistication. For this reason, the DC is helpful for screening alternative flowsheets in Stages 2 and 3 of the design process, before it is practical to consider the details of the individual controllers. Even though perfect control is assumed, the values of the steady-state DC indicate: 1. The settling time for disturbance rejection. Note that disturbance directions for which the steady-state DC is high are those for which disturbance recovery is sluggish, regardless of the sophistication of the controller. 2. The limitations due to actuator constraints. Disturbance directions for which the steadystate DC exceeds the actuator constraints are those in which offset is incurred because of actuator saturation. Assuming that the process model has been scaled such that inputs are constrained to lie within u ≤ 1 , steady-state DC values above unity indicate that the – 12S-23– actuator constraints are exceeded, and hence, such flowsheets should be avoided or modified to ensure adequate regulation. Furthermore, by observing the DC variation at higher frequencies (e.g., at the closed-loop bandwidth specified), the disturbance directions are identified for which the high-frequency modes are attenuated with difficulty or not at all. The next example shows the utility of the disturbance cost for predicting the ease of rejecting disturbances, as applied to the operation of a distillation tower. Example 12S.6 Resiliency Analysis of the “Shell Process.” To test alternative control strategies, Prett and Morari (1986) provide a linearized model, referred to as the “Shell Process,” of a distillation tower to separate crude oil into fractions in a refinery. Part of the model describes the dynamics of the two top compositions as a function of the manipulated variables (the two top draw rates) and two key disturbances (the heat removal loads in pumparound streams used to remove heat and create intermediate reflux). For this example, it is sufficient to examine the matrices specific to the nominal model: 4.05 e − 27 s P{s} = 505.s39+1 −18 s e 50 s +1 1.77 − 28 s e 60 s +1 5.72 −14 s e 60 s +1 1.2 e − 27 s P d {s} = 451.s52+1 −15 s e 25 s +1 1.44 − 27 s e 40 s +1 1.83 −15 s e 20 s +1 (12S.47) The time units in this model are minutes, and both manipulated variables and disturbances are in the range ±0.5. After scaling, the inputs (both disturbances and manipulated variables) are in the range ±1. First, the disturbance cost at steady state is computed for various disturbance vectors. For d = [1,−1]T , arg(d ) = −45° : u{0} 2 −1 4.05 1.77 1.2 1.44 1 = 5.39 5.72 1.52 1.83 − 1 and for three other disturbance directions, – 12S-24– = 0.0606 2 (12S.48) arg(d ) = 0° : u{0} 2 −1 4.05 1.77 1.2 1.44 1 = 5.39 5.72 1.52 1.83 0 = 0.3072 2 −1 4.05 1.77 1.2 1.44 1 arg(d ) = 45° : u{0} 2 = 5.39 5.72 1.52 1.83 1 = 0.6748 (12S.49) 2 −1 4.05 1.77 1.2 1.44 0 arg(d ) = 90° : u{0} 2 = 5.39 5.72 1.52 1.83 1 = 0.3676 2 Clearly, the worst disturbance to reject is d = [1, 1]T (45° or −135°), whereas d = [1, −1]T (−45° or 135°) is the easiest to overcome. Figure 12S.10 Closed-loop response of the Shell process to different disturbance directions: Solid line = y1, dotted line = y2, dashed line == u1, dashed-dotted line = u2. These observations are verified by closed-loop simulations. The RGA for the matrix P{0}, with λ11 = 1.7, indicates that the control loops can operate in a stable fashion only by pairing the inputs and outputs diagonally. Thus, diagonally paired PI controllers are tuned according to the improved IMC-based tuning rules (see Section 21.4), with KC1 = 0.29, τI 1 = 64 min, KC2 = 0.42 and τI 2 = 67 min. As shown in Figure 12S.10, simulations confirm that disturbances acting in opposite directions, d = [1, −1]T , are the easiest to reject, whereas those in the same direction, d = [1, 1]T, are the most difficult. – 12S-25– To identify potential bandwidth limitations, the DC values are computed as a function of the frequency. For two disturbances, contours of DC values, computed separately for each manipulated variable, are displayed as a function of the disturbance direction (in degrees) and frequency, as shown in Figure 12S.11. This confirms that the worst disturbance direction is d = [1, 1]T (i.e., 45°), where the two manipulated variables are highest, with u1 having relatively high steady-state values, but lower than unity, while u2 has low values at steady state, but values that exceed unity at frequencies greater than 10-1.5 = 0.03 rad/min. Consequently, the fast modes in disturbance vectors entering in this direction (45°) are not attenuated, even if perfect control were possible. Thus, the fastest settling time possible for this disturbance is approximately five times the inverse of the bandwidth limit, where DC =1, that is, 150 min. This analysis is corroborated by the responses in Figure 12S.10, which indicate that the most severe bandwidth limitations are exhibited for disturbances aligned at 45°, but with perfect steady-state disturbance rejection, with most of the static effects eliminated using u1. Clearly, the response obtained with the decentralized PI control system for d = [1, 1]T is considerably more sluggish, since delay times impose additional stability limitations, as discussed in Section 12S.4. In contrast, disturbances in the direction, d = [1, −1]T, (i.e., −45° or 135°), are more easily rejected, as the DC for each manipulated variable remains low over the entire frequency range. Again, this is corroborated by the responses in Figure 12S.10. The reader can try out this example under MATLAB, using the interactive C&R Tutorial CRGUI available on the Wiley website that accompanies this text. In the main menu, opt for the “Shell Process.” Figure 12S.11 DC contour map for the Shell Process: (a) u1; (b) u2. – 12S-26– Example 12S.7 Using DC to Improve Process Resiliency (Example 12S.3 Revisited). Returning to the process in Example 12S1.3, and noting that DRGA analysis leads to the recommendation that the variables be paired off-diagonally (i.e., y2 – u1 and y1– u2), the resiliency of the controlled system to disturbances is examined. DC contour maps for each of the manipulated variables are presented in Figure 12S.12, where it is noted that the worst disturbance is d = [20, 0]T (i.e., a disturbance direction of 0o). Furthermore, in this direction, u1 is bandwidth-limited, with saturation occurring at a frequency of approximately 10-1 = 0.1 rad/min; that is, with a characteristic time of 10 min. Consequently, the settling time in response to such a disturbance is expected to be greater than 50 min. In contrast, the second input, u2, has no bandwidth limitations. Figure 12S.12 DC contour map for Example 12S.7: (a) u1; (b) u2. The predictions afforded by the DC contour maps in Figure 12S.12 are confirmed by simulation for d = [20, 0]T. As seen in Figures 21.13(a) and (b), u1 saturates at its upper bound (with manipulated variable bounds set at | u1| ≤ 60 and | u2| ≤ 50). This bandwidth limitation is the reason for the sluggish process recovery, confirming the DC analysis that anticipates saturation in u1, a design problem that arises because the u1 range is too small to provide adequate dynamic resiliency. Through redesign, the range is increased to | u1| ≤ 70, and the performance is improved significantly, as shown in Figures 21.13(c) and (d). The reader can try out this example under MATLAB, using the interactive C&R Tutorial CRGUI available on the Wiley website that accompanies this text. In the main menu, opt for the “Mystery Process.” – 12S-27– Figure 12S.13 Closed-loop response with off-diagonal PI control to the disturbance d = [20, 0]T for Example 12S.7, with: (a,b) Original bounds on u1; (c,d) Enlarged bounds on u1 (to ±70). Shown on the top row are outputs: y1 (solid line), y2 (dashed line), and on the bottom row, inputs: u1 (solid line), u2 (dashed line). 12S.3 TOWARD AUTOMATED FLOWSHEET C&R DIAGNOSIS This section describes a procedure for assessing the controllability and resiliency of a process flowsheet that relies on heuristics to create a linearized dynamic model of the process using the results of steady-state simulations. The derived model is used to test the flowsheet C&R, using the measures introduced in Section 12S.2. As a tutorial exercise, the procedure is applied to screen the designs for the heat-integrated distillation columns in Example 12.2. Subsequently, in Section 12S.5, case studies are presented for three additional processes. For each case, the results of the approximate linear analysis are compared with the results of closed-loop simulations using nonlinear dynamic models. As will be shown, the overall approach is very promising as a short-cut diagnostic and screening tool, which can be expected to be integrated into commercial simulation software. Short-Cut C&R Diagnosis As discussed above, both steady-state and dynamic C&R analyses provide useful information for flowsheet assessment. Clearly, the second alternative provides more information – 12S-28– and is more reliable. On the other hand, steady-state analysis requires much less work, and is often adequate for screening purposes. Consequently, both approaches are considered in this section. The following steps are involved in steady-state analysis: 1. After the flowsheet is synthesized, the control structure is considered, first by selecting the process outputs to be controlled, y{t}, the manipulated variables, u{t}, and the disturbance variables, d{t}. These are related by Eq. (12S.1). 2. Steady-state simulation of the flowsheet is carried out using a process simulator. 3. Steady-state gains for the overall transfer functions, P{0} and P d {0}, are computed by perturbing each input, one at a time. 4. Steady-state C&R measures are computed using P{0} and P d {0}. For the dynamic C&R analysis (Weitz and Lewin, 1996), the steps in the algorithm are as follows: 1. Step 1 of the steady-state algorithm. 2. Step 2 of the steady-state algorithm. 3. The flowsheet is decomposed into component parts. These are MIMO subsections of the flowsheet that are approximated by matrices of low-order transfer functions (usually first order with dead time). This decomposition permits process units to be modeled in sufficient detail, allowing inverse response and overshoot phenomena to be represented. 4. Steady-state gains for the component parts are computed by perturbation of each input, one at a time. 5. Time constants and delay times are estimated assuming perfect mixing or plug flow, as appropriate, with the flow rates at steady state. At this point, transfer function matrices are defined for each component part. 6. The transfer function matrices, P{s} and P d {s} , are generated for the complete flowsheet. This involves computing the frequency response of each component part, and recombining the component parts, as dictated by the plant topology. 7. The frequency-dependent C&R measures are computed using the approximate linear model, P{ jω } and P d { jω }. – 12S-29– Many packages are available for steady-state simulation, as discussed in Chapter 5. To manipulate the linearized models in the Laplace, frequency, and time domains, MATLAB and SIMULINK are used commonly, and example scripts are introduced in Section 12S.6. The most recent commercial packages permit steady-state and dynamic simulations. These include ASPEN HYSYS, UNISIM and Aspen Dynamics, with the former used in Sections 12S.3 and 12S.5. All of the steps in the two algorithms are implemented using an array of computer packages that are becoming more integrated. Note that steps 4 and 5 deserve special attention, as they are the basis for the approximate models generated for the dynamic C&R analysis. Generating Low-Order Dynamic Models The linearized model for each component part, to be completed in steps 4 and 5 of the dynamic C&R analysis, has the form y {s} = K ⊗ Ψ {s}⋅ u {s} c c c c (12S.50) where uc{s} and yc{s} are m-dimensional input and n-dimensional output vectors, in complex space, K c is a matrix of steady-state gains in n× m real space, and Ψ {s} is a matrix describing the c dynamics in n×m complex space (each element of which is typically a delayed, low-order transfer function with dead time). The term ⊗ denotes the Schur (or element-by-element) product. Each distillation column is characterized by a single time constant. For heat exchangers, separate time constants are associated with the tube- and shell-side fluids. In the subsections that follow, it is shown that the gains, time constants, and dead times can be estimated almost entirely using the results of steady-state simulations. Steady-State Gain Matrix, K c The steady-state gains between the inputs and outputs for each component part are generated using the following procedure: 1. The material and energy balances, in the steady state, are solved for the complete flowsheet at the nominal operating point. 2. Small positive and negative perturbations are introduced for each input of each component part, one at a time, and the changes in the outputs are computed. – 12S-30– 3. The steady-state gains for each component part are computed using finite differences: Kcij = ∆yci/∆ucj, where the perturbation, ∆ucj, is sufficiently small to avoid precision losses. Dynamics Matrix, Ψ c {s} In this section, an approach is suggested for estimating the time constants and delay times for distillation columns and heat exchangers. Distillation Columns. Time constants. Following Skogestad (1987), the dominant time constant is estimated as τ = τ I + τC + τ R (12S.51) where τC and τR are the time constants (in minutes) associated with the condenser and reboiler, respectively, and τI is the time constant (in minutes) for the column, estimated according to N Mi i =1 Li τ I =∑ (12S.52) where Mi is the volumetric holdup (m3) on tray i, Li is the liquid flow rate (m3/min) from tray i, and N is the number of trays. The liquid holdup is expressed as M i = Ac (hw + how ) = πDc2 4 (hw + how ) (12S.53) where Dc is the column diameter (m), and hw and how are the weir height and fluid height above the weir (m), respectively. The latter can be expressed in terms of the weir length, lw (m), using the Francis weir equation: how Li = 111 ⋅ l w 23 (12S.54) Delay times. When the internal liquid flow rate in the column changes, a delay time is associated with the change in the fluid holdup above the weir. For a single tray, this is estimated by considering the time taken for how to stabilize after a change in the liquid flow rate (Shinskey, 1984): = θ dM ow dt dhow πDc2 = A= c 0.5 dLi dt dLi 666 ⋅ lw ⋅ how (12S.55) Thus, the overall delay experienced by the bottoms product after changes in the flow rates, temperatures, or compositions of the feed or reflux depends on the number of trays involved. In – 12S-31– contrast, it is noted that the distillate composition responds immediately to a change in the reflux flow rate, but experiences a considerable delay after changes in the feed concentration or temperature. For the latter, the delay time is estimated as the sum of the residence times on all trays between the feed and the top tray, since such a change is assumed to propagate by affecting the entire tray holdup rather than just the over-weir fluid. Typical design parameters. The following heuristics are in common use: τC = τR = 0.5τI , lw = 0.65Dc and hw = 2 in. Heat Exchangers. It is assumed that time delays associated with heat exchangers in the major processing units, such as the condensers and reboilers in distillation columns, are negligible. When heat exchangers are not included in the major processing units, they are modeled as first-order lags associated with single shell and single tube passes. Time constants. These are estimated for the tube- and shell-side fluids using τT = VT /qT and τS = VS /qS. The volumes of the fluid holdups in the tubes and shell, VT and VS, are estimated using the heat transfer area, the average fluid velocity in the tubes, v, and the tube and shell diameters. The volumetric flow rates, qT and qS, are estimated by the process simulator. Example 12S.8: C&R Analysis for Heat-Integrated Distillation Columns (Example 12.2 Revisited) Dynamic C&R analysis is applied to screen the heat-integrated distillation configurations for the dehydration of methanol in Figure 12.2 of Example 12.2. Of the three heat-integrated designs, the FS and LSR configurations provide the maximum energy savings. Clearly, the most controllable and resilient of the two should be selected based on C&R screening. Note that Chiang and Luyben (1988) prepared nonlinear dynamic models of the three heat-integrated configurations. They carried out C&R analysis, using the RGA and minimum singular values, based on linear approximations to their dynamic models. Although their findings using linear analysis were inconclusive, they showed the FS configuration to be far less desirable using closed-loop simulations with their nonlinear models. – 12S-32– In the following, each step of dynamic C&R analysis is described as it is applied to the LSF configuration. Step 1: Selection of the outputs, manipulated variables, and disturbances. As shown in Figure 12S.14, for the LSF configuration, the process outputs are the mole fractions of methanol in the three product streams (xDH, xDL and xBL). The process inputs are the control variables (LH , LL and QRH), and the disturbances are F and xF. Figure 12S.14 Component parts for the LSF configuration. Table 12S.1 Results from the Steady-state Simulation using PRO/II of Simulation Sciences for the Heatintegrated Configurations for the Dehydration of Methanol, Compared with a Single Column. Variable F, feed flow (kmol/min) xF, feed mole frac. (CH3OH) D, distillate flow (kmol/min) xD, distillate mole frac (CH3OH) B, bottoms flow (kmol/min) xB, bottoms mole frac. (CH3OH) N, number of trays NF, feed tray (1≡top) R, reflux ratio P, working pressure (mmHg) QR, reboiler duty (106kcal/min) QC, condenser duty (106kcal/min) TR, reboiler temperature (°C) TC, condenser temperature (°C) DC , column diameter (m) SC COL1 45.00 0.50 22.50 0.96 22.50 0.04 13 9 0.82 760 0.353 0.348 93.7 65.1 3.2 FS COL1 22.04 0.50 11.02 0.96 11.02 0.04 16 12 1.12 3,900 0.205 0.180 146.3 113.4 1.3 COL2 22.96 0.50 11.48 0.96 11.48 0 04 13 9 0.82 760 0.180 0.178 93.7 65.1 2.3 – 12S-33– LSF COL1 COL2 45.00 33.95 0.50 0.35 11.05 11.45 0.96 0.96 33.95 22.50 0.35 0.04 16 13 13 11 1.06 1.10 3900 760 0.222 0.175 0.175 0.205 126.5 93.7 113.4 65.1 2.0 2.4 LSR COL1 COL2 45.00 32.96 0.50 0.33 12.04 10.46 0.96 0.96 32.96 22.50 0.33 0.04 13 16 11 12 0.75 1.15 760 3,900 0.180 0.205 0.179 0.180 77.2 127.2 65.1 95.9 2.3 2.0 Figure 12S.15 Information flows between the component parts of the heat-integrated distillation configurations: (a) FS; (b) LSF; (c) LSR. Step 2: Steady-state simulation. The simulation was carried out using the PRO/II simulator and assuming that there is no pressure drop in the columns, no heat losses to the surroundings, and tray – 12S-34– efficiencies of 75%. The thermodynamic properties were computed using the UNIFAC option. These conditions were used by Chiang and Luyben (1988), with the exception that they accounted for heat losses. The results for the four flowsheets are in Table 12S.1. Note that the energy requirements for the LSR and FS configurations are the lowest (0.205×106 kcal/min), followed by the LSF configuration (0.222×106 kcal/min). Step 3: Decomposition into component parts. It has been demonstrated that a first-order lag is a reasonable approximation for the dynamics of a distillation column (Skogestad, 1987). Thus, the LSF configuration is decomposed into two component parts, one for each column. Four intermediate variables are identified to model the information transfer between the component parts: xBH, BH, TBH and QCH (= −QRL). Note that both TBH and QCH are needed for the energy balance in the reboiler because partial vaporization occurs. The control variables, in perturbation variable form, are scaled between zero and their nominal values, and the disturbances are scaled using bounds 20% above and below their nominal values. The outputs are scaled to provide a reasonable match with the steady-state gains computed by Chiang and Luyben (1988), it being noted that the output scaling does not affect the RGA or the DC values. Steps 4 and 5: Computing Kc and ψc(s). These are computed following the procedure in the section on “Generating Low-order Dynamic Models,” which gives linearized models for the high-pressure column in the LSF configuration: x DH 0.017 x 0.011e−1.3s BH 1 TBH = − 0.33e−1.3 s 13s +1 0.916e−1.3s BH 4e − 5e−1.3s QCH and for the low-pressure column: x BL 1 0.792e− 0.1s x = 17 s +1 0.790e− 8.5 s DL −1.109 0.001 −1.859 0.006 e− 0.1s − 0.2 e− 0.1s 59.0 −123.7 1.127 e− 0.1s − 0.994 0.001e− 0.1s − 0.029e− 0.1s − 0.051 0.007 e− 0.1s 0.003 – 12S-35– 0.090 e− 6.4 s LH 1.296 e− 0.1s Q RH − 41.05e− 0.1s F − 0.02e− 0.1s 0.003e− 0.1s xF − 2.161 − 3.291 x BH T BH − . s 1 4 0.012e B H 0.038 QRL LL (12S.56) (12S.57) Step 6: Generation of transfer function matrices. The linear approximation for the flowsheet dynamics is obtained by recombining the models for the component parts. Note that Figure 12S.15 shows schematically how the linearized models for the component parts are linked in each of the configurations. The inputs and outputs associated with each of the configurations are represented by the terminal junctions to the left and right. Thus, for example, the FS configuration has four manipulated and two disturbance variables (six inputs in all), and four output variables. The blocks marked HPC and LPC represent the component parts for the high- and low-pressure distillation columns, respectively. The arcs represent the flow of information (intermediate variables) to and from the component parts. The recombination to form overall transfer functions is accomplished by algebraic manipulation. For the LSF configuration, Eqs. (12S.56) and (12S.57) are rewritten in block-matrix form: x DH x BH PH (1,1) TBH = PH ( 2 ,1) B H QCH LH PH (1,2 ) Q RH PH ( 2 ,2 ) F xF (12S.58) x BH T BH x BL x = PL (1,1) PL (1,2 ) B H (12S.59) DL QRL LL where the matrix blocks contain elements from the transfer function matrices in Eqs. (12S.56) and [ ] (12S.57); for example, PH (1,1) = 1 13s +1 [0.017 − 1.109] Next, by algebraic manipulation, the vector of internal variables, namely [xBH TBH BH QCH]T, is eliminated, leading to LH Q + P (1,2 ) F 1 , 1 0 x = P ( ) [ DH ] H xF LRH H L – 12S-36– (12S.60) L x BL = P 1,1 ⋅ P 2,1 P 1,2 Q H + P 1,1 ⋅ P 2,2 F ( ) H ( ) x ( ) ( ) ( ) H L x DL L F LRH L L (12S.61) Note that Eqs. (12S.60) and (12S.61) are in the standard transfer function form of Eq. (12S.1). Similar manipulations are used for the other two configurations. The LSR configuration involves the most complicated manipulations, since it involves the feedback of information. These models are compared with those derived by Chiang and Luyben (1988), who fitted linear transfer functions to the transient open-loop responses that were computed using their nonlinear model. In Figure 12S.16, the diagonal RGA matrix coefficients for all four configurations are plotted against the frequency; the values reported by Chiang and Luyben appear on the right, while those computed using the linear models derived using the C&R analysis appear on the left. As shown, the results are in close agreement. The resonant peaks computed by Chiang and Luyben are the result of differences in the time constants and delay times in the transfer function elements. Furthermore, the relative gains, computed using the procedure in this section, do not vary significantly with frequency. Hence, diagonal pairings are preferred for the decentralized control system. Step 7: Computation of C&R measures. Figure 12S.17 shows the DC contour maps computed for each of the manipulated variables associated with the configurations SC, LSR, and FS, where the ordinate is the direction of the disturbance [F, xF]T, and the abscissa is the log 10 of the frequency. Since DC values in excess of unity correspond to saturated manipulated variables, it is apparent that the disturbances are rejected adequately by all of the designs at the steady state (i.e., when ω = 0). However, for a wide range of disturbance directions, the FS configuration has disturbance costs in excess of unity at frequencies beyond 0.1 rad/min in three of the manipulated variables (LH, QRH and FH/FL). Thus, disturbance rejection is expected to be very sluggish for this configuration. The other two configurations have low disturbance costs, and are expected to reject these disturbances nearly as well as a single column. Thus, the FS configuration should be rejected and the LSR configuration selected, because its energy requirements are lower than for the LSF and SC configurations. – 12S-37– Figure 12S.16 Diagonal RGA elements as a function of frequency for the four configurations to dehydrate methanol: (a) Procedure in Section 12S.3; (b) Chiang and Luyben (1988). To confirm these results, nonlinear dynamic simulations are carried out using HYSYS. Three configurations are simulated: (1) the single column (SC) in the LV configuration; (2) the FS configuration with the pairing: xDH –LH, xBH – QRH, xDL –LL and xBL –FH/FL ; (3) the LSR – 12S-38– configuration with the pairing: xDH –LH, xBH – QRH and xDL –LL. These pairings are selected on the basis of the RGA. The control loops tuned using the IMC-PI tuning rules (see Section 12S.4), with tuning parameters summarized in Table 12S.2. Note that the nominal values of the manipulated variables are at the mid-point of their ranges. Figure 12S.17 DC contour maps for the SC, FS and LSR configurations to dehydrate methanol. The bounds on the disturbances are ±20% from their nominal values. The DC contour maps for each manipulated variable are computed separately, with bold solid lines indicating DC = 1. See Figure 12S.39 for the DC contour maps for the LSF configuration. – 12S-39– Table 12S.2: Tuning parameters for the SC, LSR and FS Configurations. SC† Loop LSR xDL –LL Kc = 29; τI = 10 min xBL – QRL Kc = 6; τI = 10 min Kc = 1; τI = 10 min xBL – FH/FL‡ FS Kc = 1; τI = 5 min Kc = 0.5; τI = 30 min xDH –LH Kc = 0.15; τI = 5 min Kc = 1; τI = 5 min xBH – QRH Kc = 1; τI = 10 min Kc = 0.1; τI = 10 min Notes: † For the SC configuration, the temperatures on trays 2 and 10 are regulated instead of the compositions. ‡ The xBL composition controller is the master of a lower-level flow controller to regulate FH/FL. Figures 21.18, 21.19 and 21.20 show the responses for the three configurations, subjected to the worst-case disturbance in which the feed flow rate and composition simultaneously undergo positive step changes to their design limits. As shown in Figure 12S.18, the SC configuration is returned to its set points in approximately 100 min, with T10 most affected. Note that this response is qualitatively similar to that of the linear approximation shown in Figure 12S.3. The simulation can be reproduced using the METH_SC.HSC file available on the Wiley web site that accompanies this book. As shown in Figure 12S.19, the response of the FS configuration to the same disturbance is very sluggish, settling in about 100 minutes, and exhibiting severe undershoots in two of the four mole fractions: xBH (by 20 mol %) and xBL (by 10 mol %). This verifies the predictions of the DC contour maps in Figure 12S.17, which anticipate significant bandwidth limitations. These simulation results can be reproduced using the METH_FS.HSC file, also available on the Wiley web site. – 12S-40– Figure 12S.18 Response of the SC configuration to simultaneous disturbances in F (from 2,700 to 3,000 kmol/h) and xF (from 0.5 to 0.6 methanol mol fraction): T2 (dashed line), T10 (solid line), set points (dotted lines). Figure 12S.19 Response of the FS configuration to simultaneous disturbances in F (from 2,700 to 3,000 kmol/h) and xF (from 0.5 to 0.6 methanol mol fraction): xBH (solid line), xBL (dashed line), xDL (dash-dotted line), set points (dotted line). Figure 12S.20 Response of the LSR configuration to simultaneous disturbances in F (from 2,700 to 3,000 kmol/h) and xF (from 0.5 to 0.6 methanol mol fraction): xDH (dashed line), xBH (solid line), xDL (dash-dotted line), and set points (dotted line). – 12S-41– In contrast, the response of the LSR configuration to the same disturbance, shown in Figure 12S.20, settles in about half the time, with significantly less undershoot in xDL. This is because the controllers are significantly less bandwidth-limited, as predicted by the DC contour maps in Figure 12S.17. Furthermore, the settling time of the LSR configuration is comparable to that for the single column, as predicted by the DC analysis. These simulation results can be reproduced using the METH_LSR.HSC file on available on the Wiley web site. It should be emphasized that the DC contour maps are based on the assumption of perfect control, assuming that there are no stability limitations to increases in the controller gain. In practice, when single-loop controllers are implemented, the controller gains are limited, as in this example, by process interactions, or by single-loop stability limitations such as delay times. As a result, the bandwidth limitations are usually underestimated by the DC contour maps, but usually not sufficiently to affect their diagnoses when used to screen alternative designs. Note that the prediction that the FS configuration provides significantly worse disturbance rejection compared with that of the LSR configuration has been verified by simulation. Clearly, the LSR design is preferable based on energy-efficiency and controllability. This approach has been used successfully for screening more complex heat-integrated flowsheets (Weitz, 1994), exothermic reactors (Naot and Lewin, 1995), and polymerization reactors (Lewin and Bogle, 1996). In all cases, the projections were confirmed using rigorous dynamic models. To further illustrate this screening technique, Section 12S.5 provides three case studies, involving exothermic reactors in series, heat-exchanger networks, and a recycle process. – 12S-42– 12S.4 CONTROLLER LOOP DEFINITION AND TUNING Since the regulatory loops use PI controllers for verification of the C&R analysis, a brief summary of their configuration and tuning is provided in this section. Definition of PID Control Loop. This involves specifying: 1. The process variable to be controlled, PV; that is, any stream- or operation-related variable in the flowsheet (e.g., pressure, temperature, liquid level, species mass or mole fraction, mass or molar flow rate). The minimum and maximum values of the PV are used to express the PV as a percentage of its full range: PV − PVmin PV ( % ) = × 100 PVmax − PVmin (12S.62) 2. The controller output, OP, to be manipulated by the controller, as a percentage of its full range. This variable is usually either a stream flow rate or the rate of heat transfer of an energy stream. Generally, its minimum value is specified as zero and its maximum is taken as twice its nominal value. Note that occasionally the nominal value is not positioned midway between the minimum and maximum values (e.g., when the nominal flow rate of a bypass stream lies near its maximum or minimum flow rate). 3. The controller action, either direct or reverse acting, which defines the direction of its effect. For a direct-acting controller, when the PV rises above the setpoint (SP), the OP increases, and vice versa. In these cases, the static process gain is negative, as illustrated for a level controller in Figure 12S.21(a). Here, the liquid level is the PV, the flow rate of the effluent stream, Qo, is the OP, and the controller action is set to Direct. In contrast, for a reverse-acting controller, when the PV rises above the SP, the OP decreases, and vice versa. In these cases, the static process gain is positive, as illustrated in Figure 12S.21(b), which shows a different controller configuration for the surge tank. Here, the flow rate of the feed stream, Qi, is the OP, and the controller action is set to reverse. – 12S - 43 – Figure 12S.21 Level-control configurations for a surge tank: (a) Direct acting; (b) Reverse acting 4. The tuning parameters. For a PID controller, the output, OP(t), is a function of the tracking error, E(t): OP{t} = OPSS + K C E{t} + 1 τi t ∫0 E{θ}dθ + τ d dE{t} dt (12S.63) where OPSS is the bias, or controller output at zero error, and KC, τi and τd are the proportional gain, integral time constant (or reset time), and derivative time constant (or rate time) of the controller. The tracking error at time t, E(t), is the difference between the set point and the process variable: E{t} = SP{t} − PV {t} (12S.64) As mentioned above, SP, PV and OP are expressed as percentages of their full ranges. Consequently, the controller gain, KC, is dimensionless, and represents the percentage change in OP for a one-percent change in PV. In the absence of other information, factory settings are used: KC = 1, τi = 10 min. and τd = 0. These are tuned for improved performance, as discussed in the next section. Controller Tuning. The PID (Proportional-Integral-Derivative) controller is the most commonly used feedback controller in industry, with three tunable parameters as stated previously. The integral component ensures that the tracking error, E{t}, is asymptotically reduced to zero, whereas the derivative component imparts a predictive capability, potentially enhancing the performance. Despite its apparent simplicity, the subject of PID controller tuning has been discussed in several textbooks and thousands of research papers since the landmark work of Ziegler and Nichols (1942). In – 12S - 44 – practice, despite these developments, most PID controllers are tuned as PI controllers for several reasons. 1. The improved performance attainable using the derivative term is often not required. In many cases, the derivative action causes the controller to respond more nervously, especially when step changes are imposed on the set points. Furthermore, as pointed out by Luyben and coworkers (1999), it is often sufficient to tune level controllers as proportionalonly regulators (i.e., 1/τi = 0 and τd = 0). 2. The derivative component amplifies measurement noise. It is highly recommended that derivative action be applied only to filtered feedback signals. 3. Using model-based tuning methods, as discussed in the section on “Model-based PI Controller Tuning,” higher-order models are needed to tune a PID controller. In many cases, the additional engineering effort is not justified. 4. Many PID controller loops remain on their factory settings long after plant start-up. When the controller action is in the right direction and the PV range is defined wisely, these settings often give adequate performance. For tuning, either on-line methods, implemented with controllers on-line, or modelbased methods, which rely on process models, are utilized. The main advantages of the on-line methods are that tuning occurs under closed-loop control and a process model is not required. It should be recognized, however, that they provide initial controller settings that are usually improved iteratively during operation. Furthermore, typical on-line tuning rules apply strictly for a single control loop. For a multivariable control system, detuning is often necessary to prevent process interactions from introducing instability. Because no guidelines are available to modify the controller parameters when this occurs, the discussion below focuses on modelbased PI tuning rules. For more details on on-line tuning, the reader is referred to Seborg et al. (1989), Luyben (1990), and the section covering process dynamics on the multimedia on the Wiley website that accompanies this text (HYSYS → Dynamic Simulation → Tuning PI Controllers). – 12S - 45 – Model-based PI-controller Tuning. All control systems are implicitly or explicitly model based. They can be as simple as PI or PID controllers, which are implicitly based on first- or second-order lag models of the process, or as sophisticated as a set of DAEs that model the process and are solved in a nonlinear predictivecontrol algorithm. However, models approximate the true process dynamics and, when they involve nominal parameters with lower and upper bounds, are said to exhibit parametric uncertainty. Through the internal model control (IMC) theory (Morari and Zafiriou, 1989), the relationships among model reduction, model uncertainty, and closed-loop performance are well established. Both the selection of the nominal model order and its parameters (e.g., a first-order lag) and the model uncertainties (parameter ranges) limit the achievable closed-loop performance. Although space is not available to discuss these concepts further, the IMC tuning rules and their advantages are introduced and applied in the remainder of this section. The reader is referred to the text by Ogunnaike and Ray (1994) for a full exposition of model-based control. Figure 12S.22 (a) IMC and (b) Classical control structures. The IMC structure, illustrated in Figure 12S.22(a), includes the process, p{s}, the process model, ~ p{s} , and the IMC controller, q{s}. This structure is equivalent to the classical feedback structure, shown in Figure 12S.22(b), in which c{s} is the feedback controller. It is convenient to carry out design using the IMC structure, and then implement the control system using the classical feedback structure, with c{s} computed using the equation c{s} = (1 − ~ p{s}q{s})−1 q{s} (12S.65) The order of the process model determines the order of the controller, and therefore has an impact on the achievable performance of the control system. Thus, a PID controller, which is of second order, is generated on the basis of a model, ~ p{s} , of the same order. To design c(s) as a PI – 12S - 46 – −1 controller, the process model is limited to a first-order transfer function: ~ p{s} = K p (τs + 1) , in which case the IMC controller becomes q{s} = (τs + 1)(λs + 1)−1 K p −1 , where λ is the time constant of the IMC filter. Using Eq. (12S.65), the classical feedback controller is ( 1 c{s} = (1 − ~ p{s}q{s})− q{s} = (τs + 1) K p λs )−1 ≡ K C 1 + 1 τi s (12S.66) By equivalence of the terms, the IMC-PI tuning rules are τi = τ and K C = τ i K P λ . (12S.67) −1 For processes exhibiting time delay, e.g., ~ p{s} = K p e − θs (τs + 1) , Rivera et al. (1986) recommend the IMC tuning rules: θ τ i = τ + and K C = τ i K P λ . 2 (12S.68) IMC-based tuning parameters have been computed for a variety of open-loop stable transfer functions by Rivera et al. (1986), and for open-loop unstable transfer functions by Rotstein and Lewin (1991). The main advantage of the IMC tuning rules is that the time constant of the IMC filter is the only tunable parameter. This is of great practical importance because it provides guidance for detuning the controllers in the face of model uncertainty and multivariable system interactions. The value of λ should be set initially to either the desired closed-loop time constant (about one-fifth of the desired settling time), or to twice the process delay time, whichever is the greatest. Increased robustness (less oscillations) is attained by increasing its value, whereas more aggressive control action results from decreasing its value. Example 12S.9 Tuning PI Control Loops for a Binary Distillation Column (Example 12S.7 Revisited) As discussed in Example 12S.8, this column, for the dehydration of methanol, is simulated using ASPEN HYSYS. To assist in tuning the PI controllers, each control loop is placed in manual operation and a step change in the manipulated variable is applied. For a step change from 50 to 60% of the maximum reflux flow, the temperature of the second tray (used to regulate the distillate composition) is reduced from 66.7 to 66 oC, with a settling time of about 50 min. Since the PV range is defined from 25 to 125 oC, the 10% increase in OP causes a PV change of: – 12S - 47 – PV − PVmin 66 − 66.7 × 100 = ∆PV = × 100 = −0.7% 125 − 25 PVmax − PVmin Hence, the dimensionless process gain is: Kp = – 0.7/10 = – 0.07, with direct action control needed due to the negative process gain. Furthermore, the open-loop process time constant is approximately 10 min (assuming the settling time is on the order of five time constants). Thus, the IMC-based PI tuning parameters for this loop, computed using Eq. (12S.67), are: τi = τ = 10 min, and K C = τi K Pλ = 143 λ The value of λ is tunable, allowing the designer to trade-off between robustness and performance. With λ = 5 min, KC = 29. A similar approach for the bottom loop leads to the PI tuning: KC = 6, τi = 10 min. Example 12S.10 Tuning PI Control Loops for the Shell Process (Example 12S.6 Revisited) In Example 12S.6, RGA analysis for the Shell process indicates the need for diagonal pairing. For the loop, y1 – u1, the open-loop transfer function is: ~ p{s} = 4.05e −27 s (50 s + 1)−1 . Using Eq.(12S.68), with λ = 2θ , the PI tuning parameters are: KC1 = 0.29, τi1 = 64 min. Similarly, for the second loop, y2 – u2, the PI tuning parameters are: KC2 = 0.42, τi2 = 67 min. As shown in Figure 12S.10, these settings give adequate closed-loop response. For more details on the implementation and tuning of PI controllers using ASPEN HYSYS, the reader is referred to the multimedia on the Wiley website that accompanies this text (HYSYS → Dynamic Simulation → Tuning PI Controllers ). In the following case studies, C&R analysis is demonstrated, with results verified using dynamic simulations of the PI-controlled processes. 12S.5 CASE STUDIES The case studies presented in this section demonstrate the application of C&R analysis for screening the designs of three chemical processes that are representative of those encountered – 12S - 48 – during process synthesis. The case studies also show how the results using steady-state and dynamic C&R analyses compare. Case Study 12S.1 Exothermic Reactor Design for the Production of Propylene Glycol (Example 12S.1 Revisited) Returning To Step 1 of Example 12S.1, the CSTR to hydrolyze propylene oxide to propylene glycol is considered further. Its material and energy balances are in Eq. (12S.6) and (12S.7), with variables defined and specifications provided. Beginning with these dynamic balances, the steadystate behavior patterns for this reactor are examined next. Steady-State Solution Eq. (12S.8) provides an expression for the PO concentration in the reactor in terms of the reactor temperature: C PO {T } = 1 + 4k {T }C PO ,inτ − 1 2k {T }τ (12S.69) where CPO,in = ℑPO,in/(q0 + qw) and the residence time, τ = V/(q0 + qw). Note that the fractional conversion of PO is, X = 1 – CPO/CPO,in. Substituting for conversion in Eq. (12S.9), provides expressions for the heat generation and removal rates in terms of the reactor temperature: H GEN = 1 2 {T }(− ∆H ), H REM = (T − T0 ) k {T }C PO τ CP (12S.70) These are monotonically increasing functions of the reactor temperature. HGEN is small at low temperatures, where the reaction rates are small, rising exponentially with temperature to a plateau limited by the complete conversion of PO. In contrast, HREM varies linearly with the reactor temperature. Figure 12S.23 shows these rates as a function of T for qw = 5.325 m3/h. Note that a steady state occurs at the intersection of the two curves, at T = 82.4 oC, X = 0.97. For small positive perturbations in temperature in the vicinity of T = 82.4 oC, the heat removal rate is higher than the heat generation rate, whereas the opposite is true for small negative perturbations. Such imbalances result in the temperature returning to its operating point, which is referred to as stable in the open loop or open-loop stable. – 12S - 49 – Figure 12S.23 Solution diagram for PO hydration in a CSTR with qw = 5.325 m3/h: HGEN = solid line, HREM = dotted line. Figure 12S.24 Solution diagram for PO hydration in a CSTR with qw = 8 m3/h: HGEN = solid line, HREM = dotted line. As qw increases, τ decreases shifting HREM to the left, while increasing its gradient. At qw = 8 m3/h, three intersections with the HGEN curve occur, corresponding to three steady-state solutions, as shown in Figure 12S.24. The upper and lower intersections, at 62 and 25 oC, are stable in the open loop. However, for small positive perturbations in temperature in the vicinity of T = 44 oC, the heat removal rate is lower than the heat generation rate. This imbalance causes the reactor to move to the upper (stable) operating point. Similarly, small negative perturbations in temperature cause the reactor to move to the lower (stable) operating point. Therefore, the intermediate operating point is referred to as unstable in the open loop or open-loop unstable. This operating point can – 12S - 50 – only be maintained by installing a feedback control system, whose gain must be high enough to ensure closed-loop stability. The multiplicity of steady states exhibited by the CSTR may lead to hysteresis phenomena when operating the reactor. This is illustrated using ASPEN PLUS and ASPEN HYSYS on the multimedia on the Wiley website that accompanies this text. See the sections on modeling CSTRs. Returning to the steady state for qw = 5.325 m3/h, a conversion in excess of 95% is obtained with a 47-ft3 CSTR operating at 85% capacity, which is denoted as the nominal operating point. Assuming an aspect ratio of L/D = 2, the diameter for a single reactor is D1 = 3.1 ft. Consider an alternative design composed of two CSTRs in series, as shown in Figure 12S.27b. Assuming perfect level control by manipulating the effluent streams, two output variables, T1 and T2, are controlled by two manipulated variables, qw1 and qw2 (the water feed rates). Assuming operation at approximately the same temperature in each CSTR (about 80 oC), two 14-ft3 CSTRs in series provide the same conversion as for the base-case design involving a single reactor. Assuming the reactors are operated at 85% capacity and designed with the same aspect ratio (L/D = 2), the diameters are D2 = 2.08 ft. Because of the small vessel volumes in both of the alternative designs, it is not possible to estimate their costs using Eq. (22.54). C&R Diagnosis Based on safety considerations, the design involving two CSTRs is preferred, since it involves a smaller inventory of dangerous reagents. It remains to examine the C&R measures for these two designs. Beginning with a single CSTR, Eqs. (12S.6) and (12S.7) are linearized, as shown in Section 21.1, in the vicinity of the steady state to give P(s) and Pd(s) in Eq. (12S.1): − − P{s} = C T (s I − A) 1 BU and Pd {s} = C T (s I − A) 1 BD , where the scaled state-space matrices are: − 9.203 − 55.43 A= 0.1644 0.8870 − 0.0782 BU = − 0.0555 0 2.330 BD = 0.0070 0 C = [0 1] – 12S - 51 – (12S.71) These Jacobian matrices can also be computed by making small positive and negative perturbations in each input variable, one at a time. The perturbation magnitude is reduced until the magnitude of the resulting change in the outputs is insensitive to the direction of the input perturbation. The matrices are input scaled, assuming that the manipulated variables are nominally at 50% of their full range, PO feed rate disturbances (i.e., production rate changes) are limited to ±50% of the nominal value, and feed temperature disturbances are limited to ±5 oC. The time in Eq. (12S.71) is in minutes. Using the linear approximation, the DC contour map in Figure 12S.25 identifies the worst disturbance as ∆d = [+50%, 0 oC]T, that is, throughput changes. Even with this disturbance, the linear approximation indicates that there are no limitations to perfect disturbance rejection, since all of the DC values up to a frequency of 10 rad/min lie well below unity. Figure 12S.25 DC contour map for PO hydration in a single CSTR. A similar analysis for the two CSTRs in series yields the DC contour maps in Figure 12S.26. While the first reactor has disturbance rejection comparable to that for the single reactor, the control variable in the second reactor is saturated at steady-state for the worst disturbance direction (feed rate change alone). Hence, even for perfect control in the second reactor, the temperature setpoint cannot be maintained for this disturbance. However, this is mitigated by the fact that the conversion in the second reactor is small, and hence, the offset in T2 is expected to be small. – 12S - 52 – Figure 12S.26 DC contour maps for PO hydration in two CSTRs in series: (a) DC contour map for qw1; (b) DC contour map for qw 2. Table 12S.3 Controller Tuning Parameters for the Two Reactor Configurations. Single CSTR Configuration (see Figure 12S.27a) Loop FC-1 TC-1 LC-1 PV Range Kc τI, min Action Reverse Direct Direct Loop 0−100 kmol/h 0.1 1 o 20−120 C 1.5 4 0−100% 1 10 Two-CSTR Configuration (see Figure 12S.27b) PV Range Kc τI, min FC-1 TC-1 TC-2 LC-1 LC-2 0−100 kmol/h 20−120 oC 20−120 oC 0−100% 0−100% Reverse Direct Direct Direct Direct 0.1 0.5 3 1 1 – 12S - 53 – 1 2 4 5 5 Action Figure 12S.27 Control configurations for the alternative reactor configurations: (a) Single CSTR; (b) Two CSTRs in series. It should be noted that these results assume that perfect control is achievable. Thus, the true response is expected to be worse than that predicted based on the linear analysis, especially when single-loop PI controllers are implemented. Based upon the C&R analysis, two CSTRs are anticipated to provide the same disturbance rejection as obtained with a single CSTR. This suggests that the former should be selected, since it is the safest design. To confirm this conclusion, dynamic simulation of both systems is carried out using ASPEN HYSYS. Table 12S.3 summarizes the PI tuning parameters, with the control configurations shown in Figure 12S.27. The IMC-PI tuning parameters, computed using the approach described in Section 12S.4, are detuned for the level control loops, and ensure relatively tight control on the reactor temperatures. Perfect pressures control is assumed, by adjusting the small flow rates in the vapor vents (3 bar for the first vessel, and 2 bar for the second). The pressure control loops are not simulated explicitly. – 12S - 54 – Figure 12S.28 Response of the single CSTR to 50% positive and negative changes in throughput: (a) feed rate and set point; (b) reactor temperature; (c) water feed rate (%); (d) Reactor level. Because the DC analysis indicates that the most challenging regulatory control is associated with throughput changes, the organic feed rate is adjusted, and closed-loop simulations are performed to check the effects of positive and negative changes in q0. The response of the single CSTR is shown in Figure 12S.28, and indicates that both positive and negative throughput changes of 50% are easily handled, with the reactor temperature returned to its setpoint in about 15 min, while the coolant flow rate, qw, remains within its constraints, as predicted by the DC analysis. The response of the liquid level is more sluggish by design. The responses of the two-CSTR system compare well with that for the single CSTR, as shown in Figure 12S.29. Its response to a positive feed rate change is rapid, with no evidence of saturation on either manipulated variable. In contrast, – 12S - 55 – the coolant flow rate to the second reactor, qw2, saturates in response to a negative change in the throughput, as predicted by the DC analysis. However, as seen in Figure 12S.29, this does not significantly affect T2. Figure 12S.29 Response of the two-CSTR system to 50% positive and negative changes in throughput: (a) feed flow rate and set point; (b) reactor temperatures (T1 – solid T2 – dashed); (c) water feed flow rate (qw1 – solid qw2 – dashed, in %); (d) holdups (L1 – solid L2 – dashed). For more details, the reader is referred to the section covering dynamic simulation with ASPEN HYSYS on the multimedia available on the Wiley website accompanies this text. The results in Figures 12S.28 and 12S.29 can be reproduced using the files, CSTR_1.hsc and CSTR_2.hsc. – 12S - 56 – In summary, it would appear that despite indicating potential saturation problems associated with the more complex two-CSTR system, C&R analysis shows that the dynamic responses of the two systems are approximately the same. This suggests that the two-CSTR system should be adopted, since it is the safest. However, designs involving smaller hold-ups are often less resilient, especially in the face of disturbances that manifest themselves rapidly. This example has focused on the importance of performing C&R analysis in reactor design. For more examples, the reader is encouraged to study the following publications: 1. Luyben et al (1999): Chapter 4 discusses the design of control systems for reactors in general. The design of heat-integrated reactor systems is discussed in Chapter 5. 2. Shinskey (1988): Chapter 10 discusses reactor control in industrial practice. 3. Lewin and Bogle (1996): This paper concerns the optimal operation and controllability of a continuous polymerization reactor. 4. Russo and Bequette (1998): This paper discusses the multiplicity of steady-states associated with jacketed polymerization reactors. Case Study 12S.2 Two Alternative Heat Exchanger Networks (Examples 12.1 and 12.5 Revisited) Here, the two alternative heat exchanger networks (HENs) in Examples 12.1 and 12.5 are screened using C&R analysis. More specifically, the two designs are required to be resilient to ±5% changes in F1 and ±5 oF in T0. As discussed in Chapter 12, it is often necessary to augment the process degrees-of-freedom to meet control objectives, either by addition of trim utility exchangers, or by adding bypasses, as is the case here. The focus of this study is in the use of resiliency analysis to select the design configuration and to adjust its nominal operating conditions. Original HEN (No Bypass). In the network shown in Figure 12S.30, only two of the target temperatures, θ2 and θ4, are controlled by manipulation of the flow rates of the two cold streams, leaving the third target temperature, T3, uncontrolled. The energy balances for this system involve 15 variables: F1, F2, F3, T0, T1, T2, T3, θ0, θ1, θ2, θ3, θ4, Q1, Q2 and Q3, two of which, θ0 and θ1, are considered to be fixed, and two, F1 and T0, are considered to be external disturbances. – 12S - 57 – Figure 12S.30 Heat-exchanger network without bypass. Three energy balances apply for each heat exchanger. For the first heat exchanger, E-100, they are: f1{x} = Q1 − F1C p (T0 − T1 ) = 0 (12S.72) f 2 {x} = Q1 − F3C p (θ 4 − θ 3 ) = 0 (12S.73) 1 3 f 3 {x} = Q1 − U1 A1 (T0 − θ 4 ) − (T1 − θ 3 ) = 0 ln[(T0 − θ 4 ) (T1 − θ 3 )] (12S.74) For E-101, the equations are f 4 {x} = Q2 − F1C p (T1 − T2 ) = 0 (12S.75) f 5 {x} = Q2 − F2 C p (θ 2 − θ1 ) = 0 (12S.76) 1 2 f 6 {x} = Q2 − U 2 A2 (T1 − θ 2 ) − (T2 − θ1 ) = 0 ln[(T1 − θ 2 ) (T2 − θ1 )] (12S.77) Finally, for E-102, f 7 {x} = Q3 − F1C p (T2 − T3 ) = 0 (12S.78) f 8 {x} = Q3 − F3C p (θ 3 − θ 0 ) = 0 (12S.79) 1 3 f 9 {x} = Q3 − U 3 A3 (T2 − θ 3 ) − (T3 − θ 0 ) =0 ln[(T2 − θ 3 ) (T3 − θ 0 )] – 12S - 58 – (12S.80) where Ui and Ai are the heat transfer coefficients and heat transfer areas for exchanger i, respectively, such that: U1A1 = 0.0811 MMBtu/h oF, U2A2 = 0.3162 MMBtu/h oF, and U3A3 = 0.1386 MMBtu/h oF. The number of independent manipulated variables is NManipulated = NVariables − NExternally Defined − NEquations = 15 − 4 − 9 = 2, and the pairings can be selected using the RGA. To accomplish this, a linearized model is generated using the following procedure: 1. The nonlinear state equations, f{x} = 0, in Eqs. (12S.72)-(12S.80) are solved for the nominal values of the manipulated variables, u = [F2, F3 ]T, disturbances, d = [F1 ,T0]T, and constants θ0 and θ1, to determine 9 state variables: x = [T1, T2, T3, θ2, θ3,θ4, Q1, Q2, Q3]T . This is accomplished using an appropriate numerical method (e.g., the NewtonRaphson method). 2. The output vector, y = [θ2,θ4]T, is recomputed for small positive and negative perturbations of magnitude ∆ui to each manipulated variable, ui, one at a time, with the results stored in the vectors yp,i and yn,i, respectively. Then, column i of the steady-state gain matrix, P {0}, is computed: pji{0} = ∆uimax ⋅(yp,i,j − yn,i,j)/∆ui, j = 1,…,3. Note that a factor of ∆uimax scales the input variables such that ui≤ 1. 3. The output vector is recomputed for small positive and negative perturbations of magnitude ∆di to each disturbance variable, di, one at a time, with the results stored in the vectors yp,i and yn,i, respectively. Then, column i of the steady-state gain matrix, Pd {0}, is computed: pdji{0} = ∆dimax ⋅(yp,i,j − yn,i,j)/∆di, j = 1,…,3. The disturbance gain matrix is scaled arbitrarily relative to the inputs using the scaling ∆dmax= [5%, 5 oF]T. Since the nominal values of the manipulated variables are u = [F2, F3 ]T = [1.00, 1.00]T , the maximum perturbations are ∆umax = [1.00, 1.00]T. The resulting linearized model is: – 12S - 59 – ∆θ 2 − 58.7 − 73.3 ∆F ∆θ = ⋅ 2 + − − 7 14 112 . 4 ∆F3 ∆T3 − 14.3 − 41.6 P1{0} P {0} 2 2.83 2.23 4.92 1.89 ∆F1 2.94 ⋅ ∆T0 0.883 Pd {0} 1 Pd 2 {0} (12S.81) Note that the gains in Eq. (12S.81) are presented as the change in oF in response to a full-scale change of each input. Thus, for example, the linear model predicts a 4.92 oF increase in T3 in response to a 5% increase in F1. Using Eq. (12S.22), the steady-state RGA is: ( Λ = P1{0} ⊗ P1 −1 {0}) T 1.09 − 0.09 = − 0.09 1.09 (12S.82) The RGA indicates that the diagonal pairing shown in Figure 12.5, θ2 − F2 and θ4 − F3, provides responses that are almost perfectly decoupled. These are recommended, while the off-diagonal pairing has stability problems. This result is consistent with P1 (0), which is diagonally dominant. Next, the resiliency of the HEN is examined by computing the DC at steady-state for disturbances of ±5% in F1 and ±5 oF in T0: ∆F2 {0} ∆F1 ∆F2 {0} −1 ∆F {0} = −P1 {0}⋅ Pd 1{0}⋅ ∆T , DC = ∆F {0} 3 0 3 (12S.83) 2 The values of the two manipulated variables, computed to completely reject the effect of the disturbances on θ2 and θ4, lead to changes in the value of T3, computed by substituting Eq. (12S.83) into Eq. (12S.81): ) ( ∆F ∆T3 {0} = Pd {0} − P 2 {0}P1−1{0}⋅ Pd {0} 1 1 2 ∆T0 (12S.84) Table 12S.4 shows the changes in the control variables, ∆F2 and ∆F3 (assuming perfect control), the disturbance cost, and the resulting change in T3, computed using Eq. (12S.84) for four disturbance vectors. The results indicate that perfect disturbance rejection is achieved for θ2 and θ4 with negligible control effort. However, the uncontrolled temperature, T3, is significantly perturbed, with the worst-case disturbance being ∆F1 and ∆T0 in opposite directions. Variations of ±5% in F1 and ±5 oF in T0 lead to variations of approximately ±4 oF in T3. – 12S - 60 – Table 12S.4 Input changes and Disturbance Cost for the HEN without bypasses. ∆F1 ∆T0 ∆F2 +5% +5% 0 0 +5 oF +5 oF +5 oF 0.0253 0.0246 −0.0007 −0.0261 −5% ∆F3 0.0184 0.0447 0.0264 0.0080 DC = ||u||2 ∆T3 0.0313 0.0511 0.0264 0.0273 3.79 3.59 −0.20 −4.00 Figure 12S.31 Modified heat-exchanger network. Modified HEN (With Bypass). The PFD for the modified HEN, including a bypass around E-102 to eliminate the offsets in the third target temperature, is reproduced in Figure 12S.31. Resiliency analysis is used to determine the required bypass fraction. The energy balances involve 17 variables: F1, F2, F3, T0, T1, T2, T3, θ0, θ1 , θ2, θ3, θ 3′ ,θ4, Q1, Q2, Q3 and ϕ, two of which, θ0 and θ1, are assumed to be fixed, and two, F1 and T0, are considered to be external disturbances. The first six equations, (12S.72)(12S.77), for the HEN without bypasses apply. For heat exchanger E-102 and its bypass, the material and energy balances are: f 7 {x} = Q3 − F1C p (T2 − T3 ) = 0 (12S.85) f8 {x} = Q3 − F3C p (θ 3′ − θ 0 ) = 0 (12S.86) 1 3 – 12S - 61 – f 9 {x} = Q3 − K 3U 3 A3 (T2 − θ 3′ ) − (T3 − θ 0 ) = 0 ln[(T2 − θ 3′ ) (T3 − θ 0 )] (12S.87) f10 {x} = (1 − ϕ )θ 0 + ϕθ 3′ − θ 3 = 0 (12S.88) In Eq. (12S.87), the product U3A3 is identical to that for the network without bypasses ( = 0.1386 MM Btu/h oF). As the bypass fraction, ϕ, increases, K3 increases beyond unity, corresponding to an increase in the heat-transfer area. The number of independent manipulated variables is NManipulated = NVariables - NExternally Defined - NEquations = 17 − 4 − 10 = 3. This leaves F2, F3 and ϕ as the manipulated variables, which are paired with the controlled variables, θ2, θ4 and T3. A linearized model is generated and used to assist in the selection of an appropriate bypass fraction, ϕ. The procedure followed for the HEN without bypasses is used, parametrized by values of ϕ. Since the nominal values of the manipulated variables are u = [F2, F3, ϕ,]T = [1, 1, ϕ]T , the maximum perturbations are ∆umax = [1, 1, ϕ]T. For example, for ϕ = 0.1, the linearized model is: − 58.7 − 72.3 − 0.068 ∆F2 ∆θ 2 2.80 1.89 ∆F ∆θ = − 7.15 − 108 − 0.285 ⋅ ∆F + ⋅ 1 2 20 2 94 . . 4 3 ∆T0 − 14.3 − 44.9 0.237 ∆T3 ∆ϕ 4.95 0.88 P{0} Pd {0} (12S.89) Using P (0) in Eq. (12S.89), the steady-state RGA is computed using Eq. (12S.22): ( Λ = P{0} ⊗ P −1 {0}) T 1.17 − 0.22 0.04 = − 0.07 0.84 0.23 − 0.10 0.38 0.72 (12S.90) Hence, the diagonal pairing shown in Figure 12.6 is preferred; that is, θ2 − F2, θ4 − F3 and T3− ϕ, with significant interactions between the second and third loops anticipated. The impact of the bypass fraction on the resiliency of the HEN is examined next. The manipulated variable values and the disturbance cost are computed for disturbances of ±5% in F1 and ±5 oF in T0. Table 12S.5 shows the changes in the control variables, ∆F2, ∆F3, and ∆ϕ (assuming perfect control), and the disturbance cost, for four disturbance vectors, d = [F1+∆F1, T0+∆T0]T. Note that for the worst-case disturbance (∆F1 = − 5% and ∆T0 = +5 oF), the scaled – 12S - 62 – change in the bypass fraction is ∆ϕ = 12.3, which far exceeds unity. To avoid this, the nominal bypass fraction is increased further to account for the expected disturbance levels, noting that the heat exchanger E-102 must be resized. Table 12S.5 Input changes and Disturbance Cost for the HEN with ϕ = 0.1. ∆F1 ∆T0 ∆F2 ∆F3 ∆ϕ +5% +5% 0 −5% 0 +5 oF +5 oF +5 oF −0.0010 −0.0003 0.0007 0.0017 0.051 0.075 0.025 −0.026 −11.4 −10.3 0.98 12.3 DC = ||u||2 11.4 10.3 0.98 12.3 With the nominal bypass fractional flow increased to ϕ = 0.25, the linearized model is recomputed: − 58.7 − 69.8 − 0.720 ∆F2 ∆θ 2 2.80 1.89 ∆F1 ∆θ = − 7.15 − 97.1 − 3.02 ⋅ ∆F + ⋅ 2 . 10 2 . 94 4 3 ∆T 0 ∆ϕ ∆T3 2.52 − 14.3 − 53.7 5.03 0.88 P (0 ) Pd (0 ) (12S.91) In this case, the steady-state RGA is: ( Λ = P(0 ) ⊗ P −1 (0)) T 1.17 − 0.21 0.04 = − 0.07 0.75 0.32 − 0.10 0.46 0.64 (12S.92) This RGA is similar to that obtained with ϕ = 0.1, again indicating a diagonal pairing, as shown in Figure 12.6. Next, the resiliency is tested, with the results reported in Table 12S.6. Note that when ϕ = 0.25, the disturbance rejection is nearly acceptable, with DCmax = 1.1, only slightly above unity. Table 12S.6 Input changes and Disturbance Cost for the HEN with ϕ = 0.25. ∆F1 ∆T0 ∆F2 ∆F3 +5% +5% 0 −5% 0 +5 oF +5 oF +5 oF −0.0010 −0.0003 0.0007 0.0017 0.051 0.075 0.025 −0.026 – 12S - 63 – ∆ϕ −0.93 0.75 0.18 1.11 DC =||u||2 0.93 0.75 0.18 1.11 Clearly, the resiliency of the HEN increases with the nominal bypass fraction, but at the cost of increased heat-transfer area. Table 12S.7 shows the trade-off between resiliency and heattransfer area. Note that while only 12% additional heat-exchange area is required for ϕ = 0.1, the resiliency is inadequate. In contrast, when ϕ = 0.30, the resiliency is satisfactory (with DC significantly lower than unity), but the heat-transfer area is doubled. A good compromise is to select ϕ = 0.25, which approximates the desired resiliency, while requiring only 55% more heatexchange area. Table 12S.7 Trade-off between the heat-exchanger area and bypass fraction. ϕ DC = ||u||2 K3 0.10 0.15 0.20 0.25 0.30 12.3 4.63 2.16 1.11 0.58 1.12 1.21 1.33 1.55 2.05 The C&R analysis in the steady state predicts the superior performance of the modified HEN, which allows all three target temperatures to be controlled at their setpoints in the face of disturbances in the feed flow rate and temperature of the hot stream. More specifically, the steadystate RGA indicates that a decentralized control system can be configured for the modified HEN in which θ2 − F2, θ4 − F3 and T3− ϕ are paired, and in which the first loop is almost perfectly decoupled, with moderate coupling between the other two loops. Finally, aided by DC analysis, the nominal bypass fraction is selected to be 0.25, providing the best trade-off between increased plant costs and adequate resiliency. Given the design decision to use ϕ = 0.25, based upon the steady-state C&R analysis, verification is performed by dynamic simulations with ASPEN HYSYS. The hot stream of n-octane at 2,350 lbmol/h is cooled from 500 to 300 oF using n-decane as the coolant, with F2 = 3,070 lbmol/h and F3 = 1,200 lbmol/h. Note that these species and flow rates are chosen to match the heatcapacity flow rates defined by McAvoy (1983), with F1 slightly increased to avoid temperature crossovers in the heat exchangers due to temperature variations in the heat capacities. Additional details of the ASPEN HYSYS simulation are: – 12S - 64 – (a) The tubes and shells for the heat exchangers provide 2 min residence times. (b) The feed pressures of all three streams are set at 250 psi, with nominal pressure drops of 5 psi defined for the tubes, shells and for the bypass valve, V-3. Subsequently, these pressure drops are computed based on the equipment and valve sizing and the pressureflow relationships. (c) The bypass valve V-3 is sized carefully, ensuring that the nominal bypass fraction is 0.25, with the nominal valve position being 50% open (selecting a linear characteristic curve). (d) IMC-PI tuning parameters are presented in Table 12S.8. Table 12S.8 IMC-PI tuning parameters for the alternative HENs. HEN without bypass (Figures 12S.30 and 12.5) Loop θ2 − F2 θ4 − F3 PV Range,oF 300-500 300-500 Kc 2 1.5 τI, min 1.5 2.5 Action Direct Direct HEN with bypass (Figures 12S.31 and 12.6) Loop PV Range,oF Kc τI, min Action θ2 − F2 θ4 − F3 T3− ϕ 300-500 300-500 300-500 2 1 1 1 2 1 Direct Direct Reverse The regulatory responses of the two configurations are discussed next. Figure 12S.32 shows that, as predicted by the DC analysis, even the worst-case disturbance has little effect on the two controlled variables, whose control loops are decoupled, as indicated by the RGA analysis. Moreover, the uncontrolled output, T3, exhibits offsets of about ±4.5 oF, which compare well with the value of ±4 oF predicted by the linear DC analysis. In comparison, Figure 12S.31 shows that, for the HEN with bypass, the response also corroborates the results of the linear DC analysis. Most importantly, the design with ϕ = 0.25 rejects the worst-case disturbance with no saturation, indicating that the DC analysis is slightly conservative. In addition, the first control loop (θ2 − F2) is perfectly decoupled, with slight interactions seen in the other two loops, again as predicted by the static RGA analysis. For more details, the reader is referred to the section covering dynamic simulation using ASPEN HYSYS on the multimedia CD-ROM that accompanies this text, where – 12S - 65 – the files HEN_1.hsc and HEN_2.hsc, are provided to enable the reproduction of the results in Figures 12S.32 and 12S.33. Figure 12S.32 Response of HEN without bypass to the worst-case disturbances: (a) Normalized changes in F1 (solid) and T0 (dashed); (b) Tracking errors (θ2 – solid; θ4 – dashed; T3 – dotted); (c) Manipulated variables (F2 – solid; F3 – dashed). Figure 12S.33 Response of HEN with bypass to the worst-case disturbance: (a) Normalized changes in F1 (solid) and T0 (dashed); (b) Tracking errors (θ2 – solid; θ4 – dashed; T3 – dotted); (c) Manipulated variables (F2 – solid; F3 – dashed; V1 – dotted). – 12S - 66 – While steady-state C&R analysis often provides a good assessment of the controllability and resiliency, dynamic analysis should be considered when the steady-state analysis is inconclusive. The latter methods are discussed by Wolff et al. (1991) and Mathisen et al. (1993). Case Study 12S.3 Interaction of Design and Control in the MCB Separation Process Denn and Lavie (1982) show that recycles increase the process response time and static gain. Furthermore, when the recycle loop contains a time delay, resonant peaks comparable in magnitude to the steady-state gain may result. Since these phenomena are potentially destabilizing, control systems for recycle processes should be designed carefully. In this regard, control systems for recycle processes are designed using the nine-step design procedure of Luyben and coworkers (1999), presented in Section 12.3, with particular emphasis on the need to impose flow control on each recycle stream. Figure 12S.34 Flowsheet for the MCB separation process. – 12S - 67 – Process Description. Figure 12S.34 shows the Monochlorobenzene separation process introduced in Section 5.4. The process involves a flash vessel, V-100, an absorption column, T-100, a distillation column, T101, a reflux drum, V-101, and three utility heat exchangers. As shown in Figure 5.23, most of the HCl is removed at high purity (96 mol % by design) in the vapor effluent of T-100. However, in contrast with the “treater” to remove the residual HCl in the design shown in Chapter 5, in Figure 12S.34 this is removed in the small vapor overhead purge in T-101. The Benzene and Monochlorobenzene are obtained at high purity as distillate (99 mol % Benzene) and bottoms liquid products (98 mol % MCB) in T-101. It is required to design a control system to ensure that the process meets its quality specifications in the face of changes in the throughput demand (treated as a disturbance) and feed composition changes, as listed in Table 12S.9. A preliminary control system configuration is proposed, and then refined and checked using the C&R analysis. Finally, the performance of the control system is verified using dynamic simulation. Table 12S.9 Process disturbance scenarios. Species HCl Benzene MCB Total d1 d2 Molar flow rates in kmol/h 15 60 75 150 15 50 35 100 Nominal 10 40 50 100 Preliminary Control System Configuration. The nine-step control design procedure of Luyben and co-workers is applied to design the preliminary control structure in Figure 12S.35: Step 1. Set objectives. To achieve the primary control objective, the production level is maintained by flow control of the feed stream using valve V-1. Step 2. Define control degrees of freedom. As shown in Figure 12S.34, the process has twelve degrees of freedom with four valves controlling the flow rates of the utility streams (V-2, V-5, V-9 – 12S - 68 – and V-10), one controlling the feed flow rate (V-1), three controlling product stream flow rates (V6, V-8 and V-11), and the four remaining valves controlling internal process flow rates. Having chosen constant feed flow in Step 1, the feed valve (V-1) is reserved for independent flow control. Step 3. Establish energy management system. The steam valve, V-2 is used to control the flash feed temperature. Furthermore, the temperature of the recycle and bottoms product streams is contolled by adjusting the coolant valve, V-10. Step 4. Set the production rate. As stated previously, the feed valve, V-1, is assigned to a flow controller, whose setpoint regulates the production rate. Step 5. Control product quality, and meet safety, environmental and operational constraints. The pressure in V-100 is controlled by adjusting its vapor stream using valve, V-3. Pressure regulation in the T-101 is carried out by adjusting V-5, the coolant valve to the condenser E-101. Since both of the products from T-101 are required to meet specifications, the LV configuration is implemented, noting that the reflux ratio in the column is less than five. Thus, the reflux valve, V-7, is adjusted to control the distillate composition, and the reboiler steam valve, V-9, is used to regulate the bottoms composition. Step 6. Fix recycle flow rates and vapor and liquid inventories. The obvious choice for recycle flow control is valve V-12. The liquid inventories in the flash drum, the reflux drum and the column sump, are regulated using the valves V-4, V-8 and V-11, respectively. Note that the purge stream that removes the residual HCl from the column overheads is less that 1% of the feed by design. Thus, the valve V-6 is designed to fixed at 50% open and left uncontrolled. Regulations of the vapor inventories in both V-100 and T-101 have been addressed, by installing pressure controllers. Steps 7 and 8. Check component balances and control individual process units. The HCl in the feed is removed from the process, mostly in the T-100 overhead stream, with small traces removed in the column purge stream. The benzene and MCB fed to the process are mostly removed in the distillate and bottoms streams from T-101, respectively, with small traces removed with the HCl product and in the purge. Step 9. Optimize economics and improve dynamic controllability. It is noted that all of the control valves have been assigned, but the HCl product quality is still uncontrolled. To correct this, a – 12S - 69 – cascade controller is installed to regulate the HCl product stream composition, which adjusts the set point of either (a) the recycle flow controller, FC-2, or (b) the recycle temperature controller, TC-2. Figure 12S.35 shows the first alternative, which manipulates the liquid feed rate to the absorber to control the mass transfer of the organic species from the vapor stream. Clearly, quantitative methods are required to enable the most appropriate configuration to be selected, as will be shown next. Figure 12S.35 Control system for the MCB separation process. Control System Refinement using C&R Analysis. Controllability and resiliency analysis has two roles in the improvement of the control system in Figure 12S.35: (1) The RGA aids in defining the appropriate pairing between the controlled outputs and manipulated variables where interaction is anticipated; (2) The DC assists in checking that the operating ranges of key manipulated variables is sufficient to ensure adequate – 12S - 70 – disturbance rejection. To provide data for these two analytical methods, a dynamic simulation of the MCB separation process is developed using ASPEN HYSYS. The equipment items are sized as follows: (a) The flash vessel, V-100, condenser, V-101 and reboiler, E-102, are installed assuming at least 10 min liquid residence time, computed using the steady-state liquid feed rate as a basis. Thus, for example, since the liquid feed to V-100 is nominally 137 ft3/hr, the required vessel volume is 2×10×137/60 = 45.7 ft3, which is rounded up to 50 ft3. Similar calculations give volumes of 120 ft3 for V-101 and 240 ft3 for E-102. (b) The absorption column, T-100, is a 10-stage packed bed with a diameter of 1.5 ft. (c) The distillation column, T-101, has 10 valve-trays with a diameter of 2.5 ft. (d) The two heat exchangers are approximated as heat-requirement units, which assume that the control variable is the heat transfer duty. Thus, E-100 is installed as a heater, with volume of 20 ft3 and E-103 as a cooler, with a volume of 50 ft3. More detailed modeling is possible by using heat exchangers, allowing the manipulation of steam and cooling water flows. Pressure drops in these heat exchangers are defined by assigning a pressure-flow relationship, established automatically by ASPEN HYSYS on the basis of nominal flow rates. (e) A number of valves are installed to enable flow and pressure regulation of the process. Each valve is set to be 50% open, sized on the basis of nominal flow rates, and then assigned to follow a pressure-flow relationship. When a valve is selected to provide control, it is assigned to a controller, which manipulates the percentage valve opening. One valve that is maintained at 50% open is V-6, which is intended to purge the residual light gases in the feed to T-101. Several of the control loops in Figure 12S.35 are required to ensure inventory control, namely, all three level control loops and the two pressure control loops. Note that the pressure in V100 is assumed constant and the loop PC-1 is not simulated explicitly in the ASPEN HYSYS simulation. In contrast, as pointed out repeatedly in the literature, pressure control in the column is crucial to stabilize the internal flows in the column. Finally, the feed flow rate and temperature controllers are clearly decoupled from the rest of the process, and therefore need not be included in – 12S - 71 – the C&R analysis. Thus, the interactions that need to be analyzed are the effects of the four valves: V-7, V-9, V-10 and V-12 (or more precisely, the setpoint to FC-1), on four controlled variables: xD,2, xB,3, xA,1 (the mole fractions of the benzene in the distillate, MCB in the bottoms and HCl in the absorber overhead stream, respectively) and TR, the recycle temperature. Note that to improve dynamic performance, the temperature of tray 4 is controlled instead of the distillate benzene composition. The interaction analysis is performed using the steady-state RGA. To generate information to compute RGA, the loops under test in the simulated process are placed in “manual” mode, and the process is simulated to “line-out” the outputs at open-loop steady-state values. Then, distinct step changes in the four valve positions are imposed, and the new steady-state values of the outputs recorded. Note that for consistency, the step direction is chosen such that its effect on AC-1 is in the same direction. The results of these simulations are recorded in Table 12S.10. Thus, for example, a 0.5% increase in the position of the reflux valve (V-7) leads to a decrease of 4.5 oF in temperature in tray 4. Table 12S.10 Simulation results for RGA calculations. R (V-7) Range Before After Change Before After Change Before After Change Before After Change 0-100 % xD,2 (AC-1) o 100-300 F xB,3 (AC-2) xA,1 (AC-3) TR(TC-2) 0.5-1.0 0.5-1.0 50-250 oF 43.0 % 43.5 % +0.5% 226.3 oF 221.8 oF –4.5 oF 0.9857 0.9638 –0.0219 0.9596 0.9590 –0.0006 121.2 oF 117.9 oF –3.3 oF QR (V-9) xD,2 (AC-1) xB,3 (AC-2) xA,1 (AC-3) TR(TC-2) o 45.4 % 44.9 % –0.5% 226.3 F 221.5 oF –4.8 oF 0.9857 0.9576 –0.0281 0.9596 0.9589 –0.0007 121.2 oF 116.8 oF –4.4 oF FR (FC-2) xD,2 (AC-1) xB,3 (AC-2) xA,1 (AC-3) TR(TC-2) 45.0 % 45.5 % +0.5% 226.3 F 224.9 oF –1.4 oF 0.9857 0.9817 –0.0040 0.9596 0.9582 –0.0014 121.2 oF 122.1 oF +1.1 oF QC (V-10) xD,2 (AC-1) xB,3 (AC-2) xA,1 (AC-3) TR(TC-2) 0.9857 0.9803 –0.0054 0.9596 0.9605 +0.0011 121.2 oF 119.0 oF –2.2 oF 76.0 % 76.5 % +0.5% o o 226.3 F 224.3 oF –2.0 oF – 12S - 72 – Dimensionless static gains are computed, accounting for the full range of each variable. Thus, for example, the gain that relates variations of xD,2 to changes in R is: ∆x D ,2 − 4.5 200 (12S.93) = −4.5 ∆R 0.5 100 In the same way, the other 15 static gains are computed, given the overall steady-state transferp11{0} = = function matrix relationship: xD ,2 − 4.50 x − 8.76 B ,3 = x A,1 − 0.24 TR − 3.30 4.80 − 1.40 − 2.00 R 11.2 − 1.60 − 2.16 QR 0.28 − 0.56 0.44 FR 4.40 1.10 − 2.2 QC (12S.94) The RGA is computed from this linear model: 15.8 − 11.7 − 1.35 − 1.73 − 100 84.2 7.44 9.63 Λ= 15.7 − 11.8 − 15.1 12.1 69.8 − 59.8 9.98 − 19.0 (12S.95) The large RGA elements are indicative of significant sensitivity to model uncertainty, often related to process nonlinearities. While the RGA indicates that the pairings: xD,2 – R, xB,3 – QR, xA,1 – QC, and TR – FR, provide stable response, the large RGA elements are indicative of large interactions in the process. The above results, however, suggest a simpler control structure, in which FR is maintained constant, and QC is adjusted to control xA,1, giving the steady-state transfer-function matrix relationship: x D ,2 − 4.50 4.80 − 2.00 R x = − 8.76 11.2 − 2.16 Q B ,3 − 0.24 0.28 0.44 R QC x A ,1 (12S.96) In this case, the RGA is: − 4.79 0.107 6.69 Λ = − 5.34 6.29 0.053 0.660 − 0.499 0.839 – 12S - 73 – (12S.97) This suggests performance superior to that obtained with the original confirmation using the diagonal pairings: xD,2 – R, xB,3 – QR, and xA,1 – QC , leading to the modified control system shown in Figure 12S.36. Note in particular, that the third loop is almost decoupled, with strong interactions in the two distillation-column loops. The large RGA elements associated with the LV configuration are significantly larger than those expected in a column operating independently (compared with those computed for the SC configuration in Figure 12S.16), due to the additional positive feedback contributed by the material recycle. Figure 12S.36 Improved control system for the MCB separation process. Next, the DC is computed for typical process load changes and disturbances, presented in Table 12S.9. Two scenarios are considered: d1, a 50% increase in throughput, and d2, a composition disturbance in which all three compositions are changed. Table 12S.11 shows the open-loop effect each disturbance on the four outputs, indicating that the second disturbance has the greatest effect on the top composition in T-101. The effect of the two disturbances on the three outputs controlled by the control system in Figure 12S.36, expressed in scaled perturbation variable form, are: – 12S - 74 – − 0.0195 For disturbance 1: Pd ⋅ d 1 {0} = − 0.1420 − 0.4796 (12S.98) 0.1835 (12S.99) For disturbance 2: Pd ⋅ d 2 {0} = 0.0398 0.0336 Note that the scaled perturbation variables are computed by dividing the changes on the output variables in Table 12S.11 by their full-scale ranges. This allows the steady-state DC to be computed directly: For disturbance 1 : DC{0} = −[P{0}]−1 Pd d 1 {0} −1 − 4.50 4.80 − 2.00 − 0.0195 − 1.2559 = − 8.76 11.2 − 2.16 − 0.1420 = − 0.7941 − 0.24 0.28 0.44 − 0.4796 0.9103 (12S.100) For disturbance 2 : DC{0} = −[P{0}]−1 Pd d 1 {0} −1 − 4.50 4.80 − 2.00 0.1835 0.2999 = − 8.76 11.2 − 2.16 0.0398 = 0.2208 − 0.24 0.28 0.44 0.0336 − 0.0533 (12S.101) The linear analysis suggests that the effect of the first disturbance cannot be rejected completely, because it causes the first control variable, R, to saturate (the magnitude of the DC for this variable is greater than unity). In contrast, the linear DC analysis predicts that the second disturbance is rejected relatively easily. Table 12S.11 Data for DC calculations. (a) Disturbance 1: Increased throughput by 50 % Range Before After Change xD,2 (AC-1) xB,3 (AC-2) xA,1 (AC-3) TR(TC-2) 100-300 oF 0.5-1.0 0.5-1.0 50-250 oF 0.9857 0.9148 –0.0709 0.9596 0.7200 –0.2396 226.3 oF 222.4 oF –3.9 oF 121.2 oF 141.1 oF +39.9 oF (b) Disturbance 2: Composition change. Before After Change xD,2 (AC-1) xB,3 (AC-2) xA,1 (AC-3) TR(TC-2) 226.3 oF 263.0 oF 36.7 oF 0.9857 0.9976 0.0199 0.9596 0.9764 0.0168 121.2 oF 90.8 oF –30.4 oF – 12S - 75 – Dynamic simulation using ASPEN HYSYS is used to verify the predictions of the linear C&R analysis. The control loops shown in Figure 12S.36 are all PI controllers, with tuning parameters tuned using the IMP-PI rules, given in Table 12S.12. Note that the level controllers are loosely tuned, as in Case Study 12S.1. In contrast, the distillation column pressure controller, PC-2, is tuned to ensure tight control of this key variable. The gains on the three composition controllers, AC-1, AC-2 and AC-3, are tuned to ensure that the strong interaction between them does not lead to loss of stability, while imparting acceptable regulatory performance. Figure 12S.37 Response of the MCB separation process to a 50% increase in throughput (d1): (a) molar feed rates in kmol/hr – solid = MCB, dashed = benzene, dotted = HCl; (b) changes in product purities in % – solid = MCB, dashed = benzene, dotted = HCl; (c) manipulated variables – solid = V-9 (QR), dashed = V-7 (R), dotted = V-10 (QC); (d) product flow rates in kmol/h– solid = MCB, dashed = benzene, dotted = HCl. – 12S - 76 – Figure 12S.38 Response of the MCB separation process to composition change disturbance (d2). (a) Molar feed rates in kmol/hr, (b) changes in product purities in %, (c) manipulated variables, (d) product flow rates in kmol/h. Variables as in Figure 12S.37. The simulations shown in Figures 12S.37 and 12S.38 show that: a) The 3×3 control system, paired as suggested by the RGA, provides stable performance for both disturbances. b) Both of the disturbances are step changes in the molar feed rates in the three species. Note that the control system manipulates the draw rates needed while ensuring that the product compositions stay on specification, by the action of the level controllers (see Figures 12S.37d and 12S.38d). – 12S - 77 – c) The effects of both of the disturbances on the purities of the three products are rejected successfully, despite the prediction of the linear DC analysis (see Figures 12S.37b and 12S.38b). The control action perturbations required to reject the first disturbance are greater than for the second one, which is qualitatively in agreement with the DC analysis (see Figures 12S.37c and 12S.38c). Table 12S.12 IMC-PI tuning parameters for the MCB Separation Process (See Figure 12S.36). Loop PV Range Set point Kc τi Action TC-1 FC-2 AC-1 AC-2 150-350 oF 0-200 lbmol/h 200-300 oF 0.50-1.00 MCB 270 oF 90 lbmol/h 226.3 oF 0.98 MCB 3 1.4 5 12 2 min 0.5 min 25 min 10 min Reverse Reverse Direct Reverse AC-3 0.50-1.00 HCl 0.97 HCl 12 20 min Reverse PC-2 15-40 psia 26 psia 3 0.5 min Direct LC-1 0-100% 50% 2 30 min Direct LC-2 0-100% 50% 2 30 min Direct LC-3 0-100% 50% 2 30 min Direct This case study has shown the advantages of employing C&R analysis to assist in the design of a plant-wide control system using the procedure of Luyben and co-workers. The control configuration pairing is determined using the steady-state RGA. The disturbance rejection afforded by the process is predicted incorrectly by the linear DC analysis. This indicates that non-linear approaches should be used in general. Nonlinear controllability and resiliency analysis is an area of active research (e.g., Seferlis and Grievink, 1999; Solovyev and Lewin, 2001). 12S.6 MATLAB FOR C&R ANALYSIS MATLAB and SIMULINK are invaluable tools for the frequency- and time-domain calculations required for C&R analysis. In this section, several examples are carried out using MATLAB, it being assumed that the reader is familiar with the MATLAB syntax. The reader is referred to the multimedia CD-ROM that accompanies this text for sources of these and other useful MATLAB functions and scripts for C&R analysis. In particular, the interactive C&R – 12S - 78 – Tutorial CRGUI can be used to test three example linear processes for controllability and resiliency and simulate their closed-loop response under single-loop PI control. Example 12S.11 Computing the Dynamic RGA For the system given in Example 12S.3, the MATLAB script that generates the dynamic RGA in Figure 12S.5 is: % Example 12S.3 % This script file computes the dynamic RGA for Example 12S.3 % Define a vector of frequency values on a log scale wmin=-3;wmax=1;nw=30*fix(wmax-wmin); w=logspace(wmin,wmax,nw); s=i*w; % Data for process model kp=[2.5 5;1 -4]; tp1=[15 4;3 20]; tp2=[2 0;0 0]; thp=[5 0;0 5]; %process gain matrix %process time constant %process time constant %process delay % Compute the frequency response for each element of Pij p11=kp(1,1)./(tp1(1,1)*s+1)./(tp2(1,1)*s+1).*exp(-thp(1,1)*s); p12=kp(1,2)./(tp1(1,2)*s+1)./(tp2(1,1)*s+1).*exp(-thp(1,2)*s); p21=kp(2,1)./(tp1(2,1)*s+1)./(tp2(1,1)*s+1).*exp(-thp(2,1)*s); p22=kp(2,2)./(tp1(2,2)*s+1)./(tp2(1,1)*s+1).*exp(-thp(2,2)*s); % Compute lambda(1,1) and lambda(1,2) as functions of frequency. l11=p11.*p22./(p11.*p22-p12.*p21); lam11=sign(real(l11(1))).*abs(l11); l12=-p12.*p21./(p11.*p22-p12.*p21); lam12=sign(real(l12(1))).*abs(l12); %Plot the results figure semilogx(w,lam11,'-k',w,lam12,':k','LineWidth',2) xlabel('\omega [rad/min]','FontName','Times','FontSize',14) ylabel('DRGA','FontName','Times','FontSize',14) – 12S - 79 – As discussed in Example 12S.3, the steady-state RGA suggests diagonal pairings. However, the dynamic RGA implies that these pairings are unstable for frequencies higher than about 0.5 rad/min. Thus, anti-diagonal pairings should be used. Example 12S.12 Computing Disturbance Cost Maps Consider the component parts in the LSF configuration represented by Eqs. (12S.56) and (12S.57). In this example, the elements of the transfer function matrices are entered into MATLAB and used to compute the DC contour maps for this configuration. P{ jω } and Pd { jω } are computed for each frequency, and used to compute DC for all of the disturbance directions. By looping over all frequencies, the entire DC map is calculated, and repeated for each manipulated variable separately. Note that, as mentioned in Example 12S.7, the inputs are nominally at 50% of the full range. Here, the nominal inputs are taken as LH = LL = 11 kmol/min, QRH = 0.222×106 kcal/min, and the maximum disturbance magnitudes are taken as F=18 kmol/min and xF = 0.2 (±20% of the full range). % LFS: % % This script computes P(s) and Pd(s) for the LSF configuration, given the transfer-function matrices for the two component parts. It then uses the matrices to compute DC contours % Definition of frequency and direction vectors. n=41; i=sqrt(-1); wmin=-3; wmax=0; dw=(wmax-wmin)/(n-1); tmin=0; tmax=180; dt=(tmax-tmin)/(n-1); w=logspace(wmin,wmax,n); % Frequency vector [rad/min] ome=wmin:dw:wmax; % Frequency vector in log scale. phi=tmin:dt:tmax; % Direction vector [degrees] a=pi*phi/180; % Direction vector [radians] s=w*i; % Vector complex s tt = exp(i*a); % Computing the direction in radian coordinates dd(1:n,1:2) = [real(tt'),imag(-tt')]; % tt in cartesian coordinates z=zeros(1:n,1:n); % matrix for storing computed DC values. % Gains and delay times for the high pressure column: KH=[0.017 -1.109 0.001 0.090; 0.011 -1.859 0.006 1.296; -0.33 59.0 -0.2 -41.05; 0.916 -123.7 1.127 -0.02; 4.0e-5 -0.994 0.001 0.003]; DH=[0.0 0.0 0.0 6.4; 1.3 0.0 0.1 0.1; 1.3 0.0 0.1 0.1;1.3 0.0 0.1 0.1;1.3 0.0 0.1 0.1]; % Gains and delay times for the low pressure column: KL=[ 0.792 -0.029 0.007 2.161 0.012; 0.790 -0.051 0.003 3.291 0.038]; – 12S - 80 – % Note: The coefficients in the fourth column have been multiplied by -1 % since QRL = - QCH DL=[ 0.1 0.1 0.1 0.0 1.4;8.5 0.0 0.0 0.0 0.0]; for ku=1:3 for k=1:n % % % % % % end % Looping over all manipulated variables (m=3) % Looping over all frequencies. Computing the frequency response of each component part submatrix [see Eqns. (12S.56) and (12S.57)]. ph=KH.*exp(-DH*s(k))./(13*s(k)+1); pl=KL.*exp(-DL*s(k))./(17*s(k)+1); Computing P(s) and Pd(s) at the current frequency [See Eqns. (12S.60) and (12S.61)] P=[ph(:,1:4) 0 ; pl(1,1:2)*ph(2:5,1:2) pl(:,:5)]; Pd=[ph(1,3:4) ; pl(:,1:4)*ph(2:5,3:4)]; Scaling: P(:,1)=P(:,1)*11;P(:,2)=P(:,2)*0.222;P(:,3)=P(:,3)*11; Pd(:,1)=Pd(:,1)*18;Pd(:,2)=Pd(:,2)*0.2; u2 = inv(P)*Pd*dd'; % Computing DC for i_dir = 1:n % Looping over d direction 0 → 180 z(i_dir,k) = norm(u2(ku,i_dir)); end end End of frequency loop. v = [0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0]; figure cs=contour(ome,phi,z); clabel(cs) title(['Disturbance Cost for Input ',num2str(ku)]); xlabel('log(w)'); ylabel('Direction [deg]'); % End of manipulated-variable loop This script generates the DC contour maps in Figure 12S.39 for each manipulated variable separately. Note that there is no bandwidth limitation to perfect disturbance rejection in any of the control variables. – 12S - 81 – Figure 12S.39 DC contour maps for the LSF configuration to dehydrate methanol: (a) LH; (b) QRH; (c) LL. The bounds on the disturbances are ±20% from their nominal values. The DC contour maps for each manipulated variable are computed separately, with bold solid lines indicating DC = 1. See Figure 12S.17 for the DC contour maps for the SC, FS and LSR configurations. 12S.7 SUMMARY In this chapter, the methods for short-cut C&R analysis, using the results of steady-state simulations, have been described. The methods require the use of software for the solution of material and energy balances in process flowsheets (e.g., ASPEN PLUS, HYSYS.Plant) and for controllability and resiliency analysis (i.e., MATLAB). The reader is now prepared to tackle smallto medium-scale problems, and in particular, should be able to – 12S - 82 – 1. Generate a linear model of a chemical process in one of its standard forms, using either the equations expressed in a MATLAB function, or the solution of the material and energy balances computed by a process simulator. 2. Compute the frequency-dependent process transfer functions using MATLAB, given a linear model in one of its standard forms. 3. Generate the C&R measures of relative-gain array (RGA) and disturbance cost (DC), given the process transfer functions, using MATLAB. 4. Select the appropriate pairings for a decentralized control system for the process using the static and dynamic RGAs and appropriate resiliency measures. 5. Perform C&R analysis to select between alternative process configurations, given the results of process simulations. Several examples have been selected to show how the methods are used to screen alternative flowsheets in stage 2 of the design process (Table 12.1). In the first example (Section 21.3), dynamic C&R analysis enables the most resilient heat-integrated distillation configuration to be selected. In Case Study 12S.1, two designs for an exothermic reactor, involving either one or two CSTR(s) in series, show that while the latter is more economical (assuming steady-state operation), the former is more resilient to disturbances. In Case Study 12S.2, a steady-state analysis of two heat exchanger network configurations leads to the conclusion that while a design equipped with bypasses may be subject to significant constraints leading to poor resiliency, a design without them may lead to poor dynamic performance. Here, dynamic C&R analysis is crucial. Finally, Case Study 12S.3, involves a recycle processes and shows the benefits of C&R analysis in the detailed design stage (Stage 3 in Table 12.1). REFERENCES Bequette, B.W., Process Dynamics: Modeling, Analysis, and Simulation, Prentice Hall, Englewood Cliffs, NJ, (1998). Bristol, E. H., On a New Measure of Interactions for Multivariable Process Control, IEEE Trans. Auto. Control, AC-11, 133-134 (1966). Chiang, T., and W. L. Luyben, Comparison of the Dynamic Performances of Three Heat-integrated Distillation Configurations, Ind. Eng. Chem. Res., 27, 99-104 (1988). – 12S - 83 – Denn, M. M., and R. Lavie, Dynamics of Plants with Recycle, Chem. Eng. J., 24, 55-59 (1982). Lewin, D. R., A Simple Tool for Disturbance Resiliency Diagnosis and Feedforward Control Design, Comput. Chem. Eng., 20 (1), 13-25 (1996). Lewin, D. R., and D. Bogle, Controllability Analysis of an Industrial Polymerization Reactor, Comput. Chem. Eng., 20 (S), S871-S876 (1996). Lewin, D. R., W. D. Seider, J. D. Seader, E. Dassau, J. Golbert, G. Zaiats, D. Schweitzer, D. Goldberg, M. Fucci, and R. B. Nathanson, Using Process Simulators in the Chemical Engineering Curriculum – A Multimedia Guide for the Core Curriculum, Version 2.0, Multimedia CD-ROM, John Wiley, New York (2003). Luyben, W.L., Process Modeling, Simulation and Control for Chemical Engineers, 2nd ed., McGraw-Hill, New York (1990). Luyben, W. L., B. D. Tyreus, and M. L. Luyben, Plantwide Process Control, McGraw-Hill, New York (1999). Mathisen, K.W., S. Skogestad, and E. A. Wolff, Bypass Selection for Control of Heat Exchanger Networks, Comput. Chem. Eng., 16 (S), S263-S272 (1993). McAvoy, T.J. Interaction Analysis, Instrument Society of America, Research Triangle Park, NC (1983). Morari, M. Design of Resilient Processing Plants III, A General Framework for the Assessment of Dynamic Resilience, Chem. Eng. Sci., 38, 1881-1891 (1983). Morari, M., and E. Zafiriou, Robust Process Control, Prentice-Hall, Englewood Cliffs, NJ (1989). Naot, I., and D. R. Lewin, Analysis of Process Dynamics in Recycle Systems Using Steady State Flowsheeting Tools, Proc. 4th IFAC Symposium on Dynamics and Control of Chemical Reactors, Distillation Columns and Batch Processes (DYCORD'95), Helsingor, Danish Automation Society, Copenhagen (1995). Ogunnaike, B. A., and W. H. Ray, Process Dynamics, Modeling and Control, Oxford Univ. Press, New York (1994). Perkins, J. D. The Interaction Between Process Design and Process Control, Proc. IFAC Symposium on Dynamics and Control of Chemical Reactors and Distillation Columns (DYCORD'89), 195-203 (1989). Prett, D. M., and M. Morari, Shell Process Control Workshop, Butterworth, Stoneham, MA, 355360 (1986). – 12S - 84 – Rivera, D. E., S. Skogestad, and M. Morari, Internal Model Control. 4. PID Controller Design, Ind. Eng. Chem. Res., 25, 252-265 (1986). Rotstein, G. E. and D. R. Lewin, Simple PI and PID Tuning for Open Loop Unstable Systems, I. & E. C. Res., 30, 1864-1869 (1991). Russo L. P., and B. W. Bequette, Operability of Chemical Reactors: Multiplicity Behavior of a Jacketed Styrene Polymerization Reactor, Chem. Eng. Sci., 53(1), 27-45 (1998) Sandelin, P. M., K. E. Haggblom and K.V. Waller, Indirect Two-Point Control Through One-Point Control of Distillation, in J.E. Rijnsdorp, J.F. MacGregor, B.D. Tyreus, and T. Takamatsu, (eds.), Dynamics and Control of Chemical Reactors, Distillation Columns and Batch Processes, IFAC Symposia Series 1990, No. 7, Pergamon Press, Oxford, 143-148 (1990). Seborg, D. E., T. F. Edgar, and D. A. Mellichamp, Process Dynamics and Control, Wiley, New York (1989). Seferlis, P. and Grievink, J. “Plant Design Based on Economic and Static Controllability Criteria”, Proc. of the 5th Int. Conf. of Foundations of Computer-aided Process Design, 346-350, (1999). Shinskey, F. G., Distillation Control, 2nd ed., McGraw-Hill, New York, 83-89 (1984). Shinskey, F. G., Process Control Systems, 3nd ed., McGraw-Hill, New York (1988). Skogestad, S. Studies on Robust Control of Distillation Columns, Ph.D. thesis, California Institute of Technology (1987). Skogestad, S., and M. Morari. The Effect of Disturbance Directions on Closed Loop Performance, Ind. Eng. Chem. Res., 26, 2029-2035 (1987). Solovyev, B. E. and Lewin, D. R. “A Steady-state Process Resiliency Index for Non-linear Processes,” Proc. of DYCOPS'6, Jejudo Island, Korea (2001). Stephanopoulos, G., Chemical Process Control, Prentice-Hall, Englewood Cliffs, NJ (1984). Weitz, O., and D. R. Lewin, Dynamic Controllability and Resiliency Diagnosis Using Steady State Process Flowsheet Data, Comput. Chem. Eng., 20 (4), 325-335 (1996). Weitz, O., Integration of Controllability Measures into Process Design, M.Sc. thesis, Technion (1994). Wolff, E. A., K. W. Mathisen, and S. Skogestad, Dynamics and Controllability of Heat Exchanger Networks, Proc. COPE-91, 117-128 (1991). – 12S - 85 – EXERCISES 12S.1 The following RGA matrix has been obtained for a MIMO process: 0.8 Λ = −4.0 4.3 −4.6 Note that the missing elements are unavailable. If the process is to be controlled using a decentralized control system, what are the most promising pairings? a) u1−y1, u2−y2, u3−y3. b) u3−y1, u2−y2, u1−y3. c) u1−y1, u3−y2, u2−y3. d) There is not enough information to decide. 12S.2 a. Consider a two-stream blender where m1 and m2 are the mass flow rates of species 1 and 2, F is the total flow rate, and x1 is the mass fraction of species 1 in the effluent stream. Use the relative gain array to select the control loop pairings for the effluent composition: 1. x1 = 0.8 2. x1 = 0.3 b. When blending pure streams of species 1, 2, and 3 – 12S - 86 – pair the control loops for operation at F = 1, x1 = 0.1, and x2 = 0.2. 12S.3 Two liquid phases are separated using the continuous decanter shown in Figure 12S.40. The output variables, which must be controlled are: F1 , the volumetric feed rate, P1, the operating pressure, and I, the dispersion interface level in the decanter. The positions of the three control valves, m1, m2 and m3 are the manipulated variables. Figure 12S.40 Continuous separation of two liquid phases A linear model is available to describe the process: F1 P = 1 I 2.7 8.4 8.4 m1 0.38 −0.56 −0.56 m2 −0.35 m 12 0 s s 3 The equipment manufacturer has suggested the pairings F1−m1, P1−m2, and I −m3. Are these the most appropriate to use? – 12S - 87 – 2.7 8.4 8.4 Hint: 0.38 −0.56 −0.56 −0.35 12 0 s s −1 1.8 0 0.12 = 0.0023 −0.016 0.081s 0.079 −0.56 −0.081s 12S.4 In a recent publication, a decentralized control system was proposed for an experimental reactor involving heat integration between two sections of the reactor. The manipulated variables available for control are the heat duties to the two sections, Qp and Qc. The control system is intended to regulate the operating temperatures in the two sections, Tp and Tc. The authors developed a detailed nonlinear model of the process that, after linearization, gives 30 295s + 37 2 2 T p 133s + 48s + 1 133s + 48s + 1 Q p T = 33 148s + 35 Qc c 133s 2 + 48s + 1 133s 2 + 48s + 1 They tuned simple PI controllers, and found that the closed-loop response of the overall system became faster and less oscillatory when they increased the controller gains. Use the dynamic RGA to explain this observation. 12S.5 Reproduce the DC map in Figure 12S.11 for the Shell process. 12S.6 Three component parts were given by Weitz (1994) for the FS configuration of the heatintegrated distillation columns in Figure 12.2. His linearized models are, for the highpressure column, xDH 0.018 x = 1 0.047 e −1.2 s BH 11s +1 −0.001 QCH −1.471 0.003 − 7.219 0.041e −0.2 s − 0.861 0.0003 LH 0.170e −4.8 s QRH 1.449e −0.2 s F − 0.028 H xFH And for the low-pressure column, xDL 1 −1.112 x = 16 s +1 −6.745 BL 0.0185 0.048e −1.4 s and for the feed splitter (pure gain), – 12S - 88 – 0.001 0.034e −0.3 s QRL 0.168e −6.9 s LL 1.483e −0.3 s FL xFL FH 11.72 0.490 x 0 0 FH = FL − 11.72 0.510 0 xFL 0 0 FH FL 1 F . 0 xF 1 It can be assumed that all of the inputs are nominally at 50% of their full ranges. The nominal values of the inputs are taken as LH = LL = 11 kmol/min, QRH = 0.205×106 kcal/min, FH/FL = 0.49. The maximum disturbance magnitudes are taken as ∆F = 18 kmol/min and ∆xF = 0.2 (±20% of full range). Using these models, and noting the interconnections between the component parts in Figure 21.13(a), reproduce the DC contour maps in Figure 21.17 for the FS configuration. 12S.7 Two component parts were given by Weitz (1994) for the LSR configuration of the heatintegrated distillation columns in Figure 12.2. His linearized models are, for the highpressure column, x BH 1.136e− 0.2 s x = 1 0.154e− 6.3 s DH 14 s +1 −0.045 QCH −0.047 e − 0.2 s −0.027 −0.013 0.013e − 0.2 s 0.001 0.0002 0.022e −1.4 s 0.023 −0.0008 x BL −3.425 TBL −1.551 B L −0.872 LH Q RH and for the low-pressure column, xDL 0.021 x −1.4 s BL = 1 0.010e TBL 18 s +1 − 0.272e −1.4 s 0.913e −1.4 s BL 0 −1.012 −1.772 0.005e −0.1s 50.05 − 0.144e −0.1s −112.5 0.998e −0.1s LL 0.131e −8.9 s 1.297 e −0.1s QRL . − 36.96e −0.1s F −1.085e −0.1s xF It can be assumed that all of the inputs are nominally at 50% of their full ranges. The nominal values of the inputs are taken as LH = LL = 11 kmol/min, QRH = 0.205×106 kcal/min, FH/FL = 0.49. The maximum disturbance magnitudes are taken as ∆F = 18 kmol/min and ∆xF = 0.2 (±20% of full range). Using these models, and noting the interconnections between the component parts in Figure 12S.13(c), reproduce the DC contour maps in Figure 12S.17 for the LSR configuration. – 12S - 89 – 12S.8 A product P is produced by two sequential exothermic reactions, A→B→P, with an additional endothermic reaction of B leading to an unwanted product X. These reactions are carried out in a jacketed CSTR, whose material and energy balances are C A = q( C A 0 − C A ) − k1 ( T ) ⋅ C A C B = − qC B + k1 ( T ) ⋅ C A − k 2 ( T ) ⋅ C B − k 3 ( T ) ⋅ C B C P = − qC P + k 2 ( T ) ⋅ C B T = q(T0 − T ) − − U ⋅V ⋅ A T ρc P ( 1 ρc P [k1(T )∆H1 ⋅ C A + (k2 (T )∆H 2 + k3 (T )∆H 3 )CB ] − TJ ) U ⋅V ⋅ A TJ = q J TJ 0 − TJ + T − TJ ρc P with reaction rate constants −E k i ( T ) = k i 0 ⋅ exp i , i = 1,2,3 T The controlled variables are the concentration of P in the reactor effluent, CP, and the reactor temperature, T. The manipulated variables are the feed flow rate, q, and the jacket coolant flow rate, qJ. The process disturbances are the feed concentration of A, CA0, and the feed temperature, T0. Additional information is given in Table 12S.13. Table 12S.13 Process information for Exercise 12S.8. Variable CP T q qJ CA0 T0 TJ0 UVA ρcP [M] [K] [min-1] [min-1] [M] [K] [K] K ⋅l KJ ⋅ min [KJ/K] Value Variable Value 1.00 353.15 0.15 0.10 5.00 343.15 288.15 0.225 k10 k20 k30 E1 E2 E3 ∆H1 ∆H2 [min-1] [min-1] [min-1] [K] [K] [K] [KJ/mol] 1.169×1010 1.445×1011 1.689×1011 9000 9500 9800 -40 -20 1.00 ∆H3 [KJ/mol] 120 – 12S - 90 – [KJ/mol] Using the model and Table 12S.13, compute the steady-state RGA and DC. You may assume that the disturbances in the feed concentrations are limited to within ±1M and those of the feed temperature to ±5 K, and that the manipulated variables are nominally midway between their lower and upper bounds. Based on these computations, answer the following questions: a. What are the appropriate pairings to use for decentralized control? b. What is the worst possible combination of disturbances in T0 and CA0? – 12S - 91 –