A Tuning Strategy for Unconstrained SISO Model Predictive Control

Page created by Herbert Lane
 
CONTINUE READING
Ind. Eng. Chem. Res. 1997, 36, 729-746                                       729

A Tuning Strategy for Unconstrained SISO Model Predictive
Control
               Rahul Shridhar and Douglas J. Cooper*
               Chemical Engineering Department, University of Connecticut, U-222, Storrs, Connecticut 06269-3222

               This paper presents an easy-to-use and reliable tuning strategy for unconstrained SISO dynamic
               matrix control (DMC) and lays a foundation for extension to multivariable systems. The tuning
               strategy achieves set point tracking with minimal overshoot and modest manipulated input
               move sizes and is applicable to a broad class of open loop stable processes. The derivation of an
               analytical expression for the move suppression coefficient, λ, and its demonstration in a DMC
               tuning strategy is one of the significant contributions of this work. The compact form for the
               analytical expression for λ is achieved by employing a first order plus dead time (FOPDT) model
               approximation of the process dynamics. With tuning parameters computed, DMC is then
               implemented in the classical fashion using a dynamic matrix formulated from step response
               coefficients of the actual process. Just as the FOPDT model approximation has proved a valuable
               tool in tuning rules such as Cohen-Coon, ITAE, and IAE for PID implementations, the tuning
               strategy presented here is significant because it offers an analogous approach for DMC.

Introduction                                                      The derivation of an analytical expression that com-
                                                               putes λ and its demonstration in a DMC tuning strategy
  This paper details and demonstrates a tuning strategy        is one of the significant contributions of this work.
for single-input single-output (SISO), unconstrained           Table 1 illustrates the step-by-step procedure to com-
dynamic matrix control (DMC) (Cutler and Ramaker,              pute the DMC tuning parameters, including the ana-
1980, 1983) that is applicable to a wide range of open-        lytical expression that computes λ, based on a user-
loop stable processes. The DMC control law is given            specified control horizon and a given sample time.
by
                                                                  Derivation of the analytical expression for λ considers
                                                               a first order plus dead time (FOPDT) approximation of
                 j ) (ATA + λI)-1ATej
                ∆u                                    (1)      the process dynamics. It must be emphasized that the
where A is the dynamic matrix, ej is the vector of             FOPDT approximation is employed only in the deriva-
predicted errors over the next P sampling instants             tion of the analytical expression for λ. The examples
(prediction horizon), λ is the move suppression coef-          presented later in this work all use the traditional DMC
ficient, and ∆u j is the manipulated input profile com-        step response matrix of the actual process upon imple-
puted for the next M sampling instants, also called the        mentation.
control horizon. The ATA matrix, to be inverted in the            The primary benefit of a FOPDT model approximation
evaluation of the DMC control law, is referred to in this      is that it permits derivation of a compact analytical
work as the system matrix.                                     expression for computing λ. Although a FOPDT model
   Implementation of DMC with a control horizon greater        approximation does not capture all the features of some
than one manipulated input move necessitates the               higher-order processes, it often reasonably describes the
inclusion of a move suppression coefficient, λ (Maurath        process gain, overall time constant, and effective dead
et al., 1988a). This coefficient serves a dual purpose of      time of such processes (Cohen and Coon, 1953). In the
conditioning the system matrix before inversion and            past, tuning strategies based on a FOPDT model such
suppressing otherwise aggressive control action. It is         as Cohen-Coon, IAE, and ITAE have proved useful for
often used as the primary adjustable parameter to fine-        PID implementations. The tuning strategy presented
tune DMC to desirable performance. Though the dual             here is significant because it offers an analogous ap-
effect of λ has been discussed by others (e.g., Ogunnaike      proach for DMC.
(1986)), past researchers have focused most attention             The theoretical details in this paper are organized as
on the latter effect in the selection of λ.                    follows: (i) the DMC transfer function form is derived
   This paper designs an analytical expression that            and from it the existence of a gain-scaled move sup-
computes an appropriate λ by recognizing and exploiting        pression coefficient is established; (ii) an approximation
the strong correlation between the condition number of         of the scaled system matrix, ÃTÃ, is obtained from the
the system matrix, ATA, and resultant manipulated              DMC control law; (iii) formulas for the maximum and
input move sizes. The move suppression coefficient, λ,         minimum eigenvalues of the gain-scaled (ATA + λI)
which effectively modifies the condition number of the         matrix are developed and an expression for its condition
system matrix, is computed such that the condition             number is obtained; (iv) based on this expression for the
number is always bounded by a fixed low value. By              condition number, an analytical expression for comput-
holding the condition number at a low value, desirable         ing the move suppression coefficient is derived and an
DMC closed-loop performance is achieved while the              overall DMC tuning strategy is formulated; (v) a straight-
manipulated input move sizes are prevented from                forward extension of this tuning strategy to multivari-
becoming excessive.                                            able systems is proposed; and (vi) issues important to
                                                               the implementation of the tuning strategy are discussed
  * Author to whom correspondence should be addressed.         with guidelines for selection of the sample time and the
Phone: (860) 486-4092. Fax: (860) 486-2959. Email: cooper@     control horizon.
eng2.uconn.edu.                                                   Through demonstrations of several simulation ex-
                         S0888-5885(96)00428-9 CCC: $14.00    © 1997 American Chemical Society
730 Ind. Eng. Chem. Res., Vol. 36, No. 3, 1997

Table 1. DMC Tuning Strategy
1. Approximate the process dynamics with a first order plus dead time (FOPDT) model:
                                                                        -θ s
                                                               y(s) Kpe p
                                                                   )
                                                               u(s) τps + 1
2. It is desirable but not necessary to select a value for the sampling interval, T. If possible, select T as the largest value that satisfies
                                                        T e 0.1τp and T e 0.5θp
3. Calculate the discrete dead time (rounded to the next integer):
                                                                k ) θp/T + 1
4. Calculate the prediction horizon and the model horizon as the process settling time in samples (rounded to the next integer):
                                                             P ) N ) 5τp/T + k
5. Select the control horizon, M (integer, usually from 1 to 6) and calculate the move suppression coefficient:

                                              f)
                                                   {0

                                                        (
                                                     M 3.5τp
                                                    500 T
                                                             +2-
                                                                 (M - 1)
                                                                    2          )
                                                                                        M)1
                                                                                        M>1

                                                                   λ ) fK2p
6. Implement DMC using the traditional step response matrix of the actual process and the following parameters computed
   in steps 1-5:
                                          sample time, T
                                          model horizon (process settling time in samples), N
                                          prediction horizon (optimization horizon), P
                                          control horizon (number of moves), M
                                          move suppression coefficient, λ

amples that encompass a range of process characteris-
tics, the tuning strategy is validated for set point
tracking performance using different choices of the
control horizon and sampling rates. The tuning strategy
is further validated for disturbance rejection. An ex-
ample application that validates an extension to the
tuning strategy applied to multivariable systems is also
presented.

Background
   Model predictive control (MPC) has established itself
over the past decade as an industrially important form
of advanced control. Since the seminal publication of
Model Predictive Heuristic Control (later Model Algo-
rithmic Control) (Richalet et al., 1976, 1978) and
Dynamic Matrix Control (Cutler and Ramaker, 1980,
1983), MPC has gained widespread acceptance in aca-
demia and in industry. Several excellent technical
reviews of MPC recount the significant contributions in
the past decade and detail the role of MPC from an
academic perspective (Garcı́a et al., 1989; Morari and
                                                                          Figure 1. Moving horizon concept of model predictive control.
Lee, 1991; Ricker, 1991; Muske and Rawlings, 1993) and
from an industrial perspective (Richalet, 1993; Clarke,                     Dynamic matrix control is arguably the most popular
1994; Froisy, 1994; Camacho and Bordons, 1995; Qin                        MPC algorithm currently used in the chemical process
and Badgwell, 1996).                                                      industry. Qin and Badgwell (1996) reported about 600
   MPC refers to a family of controllers that employs a                   successful applications of DMC. It is not surprising why
distinctly identifiable model of the process to predict its               DMC, one of the earliest formulations of MPC, repre-
future behavior over an extended prediction horizon. A                    sents the industry’s standard today. A large part of
performance objective to be minimized is defined over                     DMC’s appeal is drawn from an intuitive use of a finite
the prediction horizon, usually as a sum of quadratic                     step response (or convolution) model of the process, a
set point tracking error and control effort terms. This                   quadratic performance objective over a finite prediction
cost function is minimized by evaluating a profile of                     horizon, and optimal manipulated input moves com-
manipulated input moves to be implemented at succes-                      puted as the solution to a least squares problem.
sive sampling instants over the control horizon. Closed-                  Because of its popularity, this work focuses on an overall
loop optimal feedback is achieved by implementing only                    tuning strategy for DMC.
the first manipulated input move and repeating the                          Another form of MPC that has rapidly gained ac-
complete sequence of steps at the subsequent sample                       ceptance in the control community is Generalized
time. This “moving horizon” concept of MPC, where the                     Predictive Control (GPC) (Clarke et al., 1987a,b). It
controller looks a finite time into the future, is il-                    differs from DMC in that it employs a controlled
lustrated in Figure 1.                                                    autoregressive and integrated moving average (CA-
Ind. Eng. Chem. Res., Vol. 36, No. 3, 1997 731
RIMA) model of the process which allows a rigorous            phisticated analysis tools and an advanced knowledge
mathematical treatment of the predictive control para-        of control concepts for their implementation. Hence,
digm. The GPC performance objective is very similar           there still exists a need for easy-to-use tuning strategies
to that of DMC but is minimized via recursion on the          for DMC.
diophantine identity (Clarke et al., 1987a,b; Lalonde and         Tuning of unconstrained SISO DMC is challenging
Cooper, 1989). Nevertheless, GPC reduces to the DMC           because of the number of adjustable parameters that
algorithm when the weighting polynomial that modifies         affect closed-loop performance. These include the
the predicted output trajectory is assumed to be unity        following: a finite prediction horizon, P; a control
(McIntosh et al., 1991). Therefore, without any loss of       horizon, M; a move suppression coefficient, λ; a model
generality, the tuning strategy proposed in this paper        horizon, N; and a sample time, T.
is directly applicable to GPC.                                    The first problem that needs to be addressed is the
   Unconstrained SISO DMC. DMC does not always                selection of an appropriate set of tuning parameters
compete with, but sometimes complements, classical            from among those available for DMC. Practical limita-
three-term PID (proportional, integral, derivative) con-      tions often restrict the availability of sample time, T,
trollers. That is, it is often implemented in advanced        as a tuning parameter (Franklin and Powell, 1980;
industrial control applications embedded in a hierarchy       A° ström and Wittenmark, 1984). The model horizon is
of control functions above a set of traditional PID loops     also not an appropriate tuning parameter since trunca-
(Prett and Garcı́a, 1988; Qin and Badgwell, 1996).            tion of the model horizon, N, misrepresents the effect
   The unconstrained SISO DMC formulation considered          of past moves in the predicted output and leads to
in this work (eq 1) does not unleash the full power of        unpredictable closed-loop performance (Lundström et
MPC. This restricted form of DMC does not allow               al., 1995).
multivariable control while satisfying multiple process           Sensitivity Study and Final Selection of DMC
and performance objectives. However, the analysis             Tuning Parameters. Based on the above discussion,
presented here provides a foundation upon which more          candidate parameters for developing a systematic tun-
advanced tuning strategies may be developed, and this         ing strategy for DMC include the prediction horizon, P,
is illustrated later in the work with an extension to a       the control horizon, M, and the move suppression
tuning strategy for multivariable DMC.                        coefficient, λ. Though this simplifies the task of sensi-
   In any event, unconstrained SISO DMC does offer            tivity analysis, the appropriate choice of these param-
some useful capabilities. For example, past comparison        eters is strongly dependent on the sample time and the
studies between unconstrained DMC and traditional PI          nature of the process.
control (e.g., Farrell and Polli, 1990) show that DMC             Over the past decade, detailed studies of DMC pa-
provides superior performance when disturbance tuning         rameters have provided a wealth of information about
differs significantly from servo tuning. DMC has also         their qualitative effects on closed-loop performance
demonstrated superior performance in the case of plant-       (Marchetti et al., 1983; Garcı́a and Morshedi, 1986;
model mismatch, except for process gain mismatch.             Downs et al., 1988; Maurath et al., 1988a). In this
   Additionally, incorporation of process knowledge in        section, a brief sensitivity study investigates the extent
the controller architecture provides DMC with anticipa-       to which various parameters affect DMC performance.
tory capabilities and facilitates control of processes with   This study is targeted toward selection of appropriate
nonminimum phase behavior and large dead times. The           tuning parameters for developing a DMC tuning strat-
form of the performance objective provides a convenient       egy.
way to balance set point tracking with control effort,            A base case process is employed to illustrate the effect
leading to an intuitive tradeoff between performance          of adjustable parameters on DMC response for a step
and robustness. Also, the DMC form considered in this         change in set point (Figures 2-4). The base case
work allows the introduction of feed-forward control in       process has the transfer function
a natural way to compensate for measurable distur-
bances.
                                                              process 1
   Challenges in SISO DMC Tuning. Tuning of
unconstrained and constrained DMC for SISO and                                               e-50s
multivariable systems has been addressed by an array                        Gp(s) )                                   (2)
of researchers. In the past, systematic trial-and-error                               (150s + 1)(25s + 1)
tuning procedures have been proposed (e.g., Cutler,
1983; Ricker, 1991). Marchetti et al. (1983) presented          Figure 2 illustrates the importance of λ in tuning
a detailed sensitivity analysis of adjustable parameters      DMC with a control horizon greater than one manipu-
and their effects on DMC performance. The method of           lated input move. The sample time is selected for this
principal component selection was presented by Maurath        base case such that the T/τp ratio is 0.1. This sample
et al. (1985, 1988b) as a method to compute an ap-            time to overall time constant ratio lies within the
propriate prediction horizon and a move suppression           recommended range for digital controllers (Seborg, 1986;
coefficient (Callaghan and Lee, 1988). To simplify DMC        Ogunnaike, 1994). A FOPDT model approximation of
tuning, Maurath et al. (1988a) also proposed the “M )         the process described by eq 2 yields a process gain, Kp
1” controller configuration of DMC.                           ) 1, an overall time constant, τp ) 157, and effective
   Other tuning strategies for DMC have concentrated          dead time, θp ) 70. The prediction horizon and model
on specific aspects such as tuning for stability (Garcı́a     horizon are fixed at 54 and represent the complete
and Morari, 1982; McIntosh, 1988; Clarke and Scat-            settling time of process 1 in samples.
tolini, 1991; Rawlings and Muske, 1993), robustness             Graph a of Figure 2 illustrates the response to a step
(Ohshima et al., 1991; Lee and Yu, 1994), and perfor-         change in set point when the control horizon is 1 (M )
mance (McIntosh et al., 1992; Hinde and Cooper, 1993,         1) and no move suppression is used (λ ) 0). Note that
1995). Although some of the above methods provide a           for M ) 1, the set point step response is sluggish and a
complete design of DMC, they also require fairly so-          λ > 0 will only further slow the process response.
732 Ind. Eng. Chem. Res., Vol. 36, No. 3, 1997

                                                                  Figure 3. Effect of prediction horizon, P, and sample time, T, on
                                                                  DMC performance for process 1 (M ) 4, λ ) 0.14).

                                                                  prediction horizon (P ) 16 with T/τp ) 0.1) or a large
                                                                  sample time (T/τp ) 0.15 with P ) 9) does not improve
                                                                  closed-loop performance significantly. However, it has
Figure 2. Importance of the move suppression coefficient, λ, in   been shown by past researchers (Garcı́a and Morari,
tuning DMC with the control horizon, M, greater than one          1982; Clarke and Scattolini, 1991; Rawlings and Muske,
manipulated input move (T ) 16, T/τp ) 0.1, P ) 54).              1993; Scokaert and Clarke, 1994) that a larger predic-
                                                                  tion horizon improves nominal stability of the closed
   Graphs b-d show the impact of λ on performance for             loop. For this reason, the prediction horizon should be
a control horizon of M ) 4. Graph b demonstrates that,            selected such that it includes the steady state effect of
with M > 1, the lack of move suppression results in               all past computed manipulated input moves. Therefore,
dramatically aggressive control effort and a significantly        the prediction horizon cannot be used as the primary
underdamped measured output response. Graph c                     DMC tuning parameter.
shows that an intermediate response can be achieved
                                                                     Figure 4 illustrates the effect of control horizon, M,
by an appropriate choice of λ. Graph d shows that a
larger move suppression coefficient (λ ) 4.0) results in          for fixed P ) 54, T/τp ) 0.15, and λ ) 0.14. Graphs a-c
a slower response. Further increasing λ can lead to an            show that increasing the control horizon from 2 to 6 does
undesirable sluggish response for most applications.              not alter the performance significantly. Actually, the
Consequently, this study shows that the choice of λ is            only noticeable effect is a slight increase in overshoot
critical to the performance achieved by DMC.                      for a larger control horizon. This is due to the additional
   Figure 3 demonstrates the interdependence of predic-           degree of freedom from a larger control horizon. This
tion horizon, P, and sample time, T. A matrix of closed-          allows more aggressive initial moves that are later
loop response results is generated for different choices          compensated for by the extra moves available.
of P and T while maintaining the control horizon, M,                 Rawlings and Muske (1993) have shown that a
and move suppression coefficient, λ, constant. When the           necessary condition to ensure nominal stability of
prediction horizon is altered, the model horizon, N, is           infinite horizon MPC is that the control horizon should
always kept long enough to ensure that the DMC step               be greater than or equal to the number of unstable
response model correctly predicts the steady state. By            modes of the system. Nevertheless, it is clear from this
eliminating the truncation errors that result when the            study that control horizon, due to its negligible effect
model horizon does not account for the process steady             on closed-loop performance, is not well suited as the
state, Figure 3 isolates the effect of prediction horizon,        primary DMC tuning parameter.
P, and sample time, T, on DMC performance.                           A conclusion from the above discussion is that the
   The trend in the Figure 3 graphs h f e f b shows               choice of prediction horizon, P, cannot be made inde-
that by reducing prediction horizon while maintaining             pendent of the sample time, T. For stability reasons, P
a constant sample time (T/τp ) 0.1), the output response          must be selected such that it includes the steady state
becomes increasingly underdamped. Also, the trend in              effect of all past manipulated input moves; i.e., it should
graphs f f e f d illustrates that by decreasing the               equal the open-loop settling time of the process in
sample time for a constant prediction horizon (P ) 9),            samples. Stability considerations also restrict the choice
the output response becomes increasingly underdamped.             of control horizon, M. Besides, the control horizon does
Thus, Figure 3 shows that the choice of P and T are               not have a significant impact on closed-loop performance
interrelated.                                                     in the presence of move suppression. Therefore, this
   Graphs f, h, and i in Figure 3 show that a large               brief study supports the opinions of other researchers
Ind. Eng. Chem. Res., Vol. 36, No. 3, 1997 733
                                                                             response coefficient of the process, and N is the process
                                                                             settling time in samples.
                                                                                The current and future manipulated input moves are
                                                                             yet undetermined and are not used in the computation
                                                                             of the predicted output profile. Therefore, eq 3 reduces
                                                                             to
                                                                                                          N-1
                                                                               ŷ(n + j) ) y0 +               ∑ (ai∆u(n + j - i)) + d(n + j)
                                                                                                          i)j+1
                                                                                                                                                              (4)

                                                                             where the term d(n + j) lumps together possible
                                                                             unmeasured disturbances and inaccuracies due to plant-
                                                                             model mismatch. From eq 4, the predicted output at
                                                                             the current instant (j ) 0) is
                                                                                                                N-1
                                                                                           ŷ(n) ) y0 +          ∑
                                                                                                                 i)1
                                                                                                                     (ai∆u(n - i)) + d(n)                     (5)

                                                                               Since the future values of d(n + j) are not available,
                                                                             an estimate is used. In the absence of any additional
                                                                             knowledge of d(n + j) over future sampling instants, the
                                                                             predicted disturbance is assumed to be equal to that
                                                                             estimated at the current time instant. Therefore
                                                                                                                                   N-1
                                                                                d(n + j) ) d(n) ) y(n) - y0 -                       ∑
                                                                                                                                    i)1
                                                                                                                                        (ai∆u(n - i))         (6)

                                                                             where y(n) is the measurement of the current output.
                                                                               A more accurate estimate of the d(n + j) is possible,
                                                                             provided the load disturbance is measured and a reliable
Figure 4. Effect of the control horizon, M, on DMC performance
for process 1 (T ) 16, T/τp ) 0.1, P ) 54, λ ) 0.14).
                                                                             load disturbance to measured output model is available.
                                                                             Using the unit step response coefficients from this
                                                                             model, an equation similar to eq 5 above can be used to
that the one candidate best suited as the final DMC
                                                                             predict the future disturbance (Asbjornsen, 1984; Muske
tuning parameter is the move suppression coefficient,
                                                                             and Rawlings, 1993). A DMC configuration that uses
λ.
                                                                             a process model augmented with a disturbance to output
                                                                             model is known as feed-forward DMC.
Analytical Framework                                                           Now, if the predicted output is to follow the set point
   In this section, the form of the DMC transfer function                    in the wake of a set point change or unmeasured
is derived from the DMC control law. The difficulty in                       disturbance, then the current and future manipulated
using this form for development of an overall DMC                            input moves in eq 3 have to be determined such that
tuning strategy is highlighted. It is then shown how
some useful information relevant to the selection of                                ysp(n + j) - ŷ(n + j) ) 0                          j ) 1, 2, ..., P      (7)
move suppression coefficient, λ, can be extracted from
the form of the DMC transfer function.                                       Substituting eq 3 in eq 7 gives
   DMC Control Law. The cornerstone of the DMC
                                                                                                  N-1

                                                                                                  ∑ a ∆u(n + j - i) - d(n) )
algorithm is a discrete convolution or step response
                                                                             ysp(n + j) - y0 -
model of the process that predicts the output (ŷ(n + j))                                         i)j+1
                                                                                                          i

j sampling instants ahead of the current time instant,
                                                                             }}
n:
                                                                                 predicted error based on past moves, e(n + j)
                                                                                                                   j

                    j                                                                                             ∑a ∆u(n + j - i)
                                                                                                                        i                      j ) 1, 2, ..., P (8)
ŷ(n + j) ) y0 +   ∑
                   i)1
                         ai∆u(n + j - i) +                                                                        i)1
              }

                                                                                                                effect of current and future
                        effect of current and
                                                                                                                  moves to be determined
                           future moves
                                           N-1

                                            ∑ ai∆u(n + j - i)
                                          i)j+1
                                                                       (3)   The terms on the left in eq 8 represent the error between
                                                                             the predicted output and the set point that will exist
                                                                             over the next P sampling instants provided no further
                                      }

                                                effect of past moves         manipulated input moves are made by the controller.
                                                                             The term on the right represents the effect of the yet
                                                                             undetermined current and future manipulated input
In eq 3, y0 is the initial condition of the measured                         moves.
output, ∆ui ) ui - ui-1 is the change in the manipulated                       Equation 8 is a system of linear equations which can
input at the ith sampling instant, ai is the ith unit step                   be represented as a matrix equation of the form
[]
734 Ind. Eng. Chem. Res., Vol. 36, No. 3, 1997

e(n + 1)                                                                           the first element of ∆u
                                                                                                         j in eq 14 to be implemented at
e(n + 2)                                                                           the current instant, can be written as
e(n + 3)                                                                                                              P
    ·
    ·               )                                                                                    ∆u(n) )     ∑
                                                                                                                     i)1
                                                                                                                         cie(n + i)              (15)
e(n + M)
    ·

 [                                                             ]
    ·
    ·                                                                              The expression for the predicted error from eq 8 can be
                                                                                   modified to eliminate y0 using eq 5:
e(n + P)
    ·
              P×1
                                                                                                                          N-1
        a1
        a2
              0
              a1
                         0
                         0
                                     · · ·            0
                                                      0
                                                                                   e(n + j) ) ysp(n + j) - y0 -           ∑ ai∆u(n + j - i) -
                                                                                                                          i)j+1
        a3    a2         a1          ·
                                         ·            0                                                                                          d(n)
                                                                       ×                                                    N-1
                                             ·
                                                      0
         ·      ·          ·
                                                                                                                            ∑
         ·      ·          ·
        a·M   aM-1
                ·        aM-2
                           ·                          a1                                    ) ysp(n + j) - ŷ(n) -              (ai+j - aj)∆u(n - j)
                                                                                                                            j)1
         ·      ·          ·                            ·                                                                                        (16)

                                             [ ]
         ·      ·          ·                            ·
        a·
         P    a·P-1      a·
                          P-2        ...              aP-M+1
                                                        ·
                                                                 P×M               Substituting eq 16 in eq 15 gives
                                                 ∆u(n)                                        P
                                                 ∆u(n + 1)                         ∆u(n) )   ∑   ci{ysp(n + i) - ŷ(n) -
                                                 ∆u(n + 2)                   (9)             i)1
                                                                                                                    N-1

                                                 ∆u(n + M - 1)         M×1                                          ∑
                                                                                                                    j)1
                                                                                                                        (ai+j - aj) ∆u(n - j)}   (17)

or in a compact matrix notation as                                                   Let ysp(n) be a weighted average of the desired set
                                                                                   point profile over P future sampling instants:
                              ej ) A∆u
                                     j                                     (10)
                                                                                                                    P
where ej is the vector of predicted errors over the next
P sampling instants, A is the dynamic matrix, and ∆u    j                                                           ∑
                                                                                                                    i)1
                                                                                                                        ciysp(n + i)
is the vector of manipulated input moves to be deter-                                                    ysp(n) )                                (18)
mined.                                                                                                                      P
   An exact solution to eq 10 is not possible since the                                                                    ∑   ci
number of equations exceeds the degrees of freedom (P                                                                      i)1
> M). Hence, the control objective is alternatively posed
as a least-squares optimization problem with a qua-                                Using eq 18 in eq 17, the latter becomes
dratic performance objective (cost function) of the form                                      P

                              j ]T[ej - A∆u
              Min J ) [ej - A∆u           j]                               (11)
                                                                                   ∆u(n) )   ∑
                                                                                             i)1
                                                                                                 ci{ysp(n) - ŷ(n)} -
               ∆u
                j
                                                                                                              P     N-1

In the unconstrained case, this minimization problem                                                         ∑
                                                                                                             i)1
                                                                                                                 ci{ ∑ (ai+j - aj)∆u(n - j)}
                                                                                                                     j)1
                                                                                                                                                 (19)
has a closed-form solution which represents the DMC
control law
                                                                                   Replacing ysp(n) - y(n) ) e(n), the error at the current
                                 T                    T                            sample time, eq 19 becomes
                         j ) (A A) A ej
                        ∆u                       -1
                                                                           (12)
                                                                                              P     N-1                                     P
   Implementation of DMC with the control law in eq
12 results in excessive control action, especially when
                                                                                   ∆u(n) +   ∑
                                                                                             i)1
                                                                                                 ci{ ∑ (ai+j - aj)∆u(n - j)} ) ∑cie(n)
                                                                                                     j)1                       i)1
                                                                                                                                                 (20)
the control horizon is greater than 1. Hence, a qua-
dratic penalty on the size of manipulated input moves                              Transformation of eq 20 to the z-domain gives the DMC
is introduced into the DMC performance objective. The                              transfer function form:

                                                                                                                  ( ( ))
modified performance objective has the form
                                                                                            u(z)                                    1
  Min J ) [ej - A∆u        T
                  j ] [ej - A∆u
                              j ] + [∆u
                                      j ] λ[∆u
                                             j ] (13)        T                     D(z) )           )
   ∆u                                                                                        e(z)                                       P

                                                                                                                              ∑
    j
                                                                                                                                  ciai+j
                                                                                                                          N-1
where λ is the move suppression coefficient. For the                                                               1
                                                                                                                        + ∑
                                                                                                                              i)1
modified performance objective the closed form solution                                                 (1 - z )
                                                                                                              -1
                                                                                                                                         - aj z-j
                                                                                                                  P              P

                                                                                                                 ∑              ∑
takes the form (Marchetti et al., 1983; Ogunnaike,                                                                        j)1
1986):                                                                                                               ci             ci
                                                                                                                 i)1            i)1

                     j ) (ATA + λI)-1ATej
                    ∆u                                                     (14)                                                                  (21)
   DMC Transfer Function. The DMC transfer func-                                   Although eq 21 provides insight into the form of               the
tion is developed by building upon prior work by Gupta                             DMC transfer function, theoretical analysis of                 the
(1987, 1993). Let ci denote the ith first row element of                           closed-loop system is very complicated (even after             the
the pseudoinverse matrix, (ATA + λI)-1AT. Then, ∆u(n),                             assumption of a FOPDT model approximation of                   the
Ind. Eng. Chem. Res., Vol. 36, No. 3, 1997 735
process). This is primarily because it is difficult to                               G(z) ) y(z)/u(z) ) KpG̃(z)          (26)
represent the elements, ci, of the pseudoinverse matrix
analytically in terms of the step response coefficients.
  Consequently, eq 21 does not serve as a convenient              Using eq 25 and eq 26, the open-loop transfer function
starting point for devising the analytical expression for         is of the form
computing λ. However, as shown in the next section,
some useful information regarding the form of the move              D(z)G(z) ) y(z)/e(z) ) (D̃(z)/Kp)KpG̃(z) ) D̃(z)G̃(z)
suppression coefficient rule can be extracted from the                                                                 (27)
DMC transfer function.
  Gain Scaling of the Move Suppression Coef-
ficient. Previous work has proposed that the choice of            and the closed-loop transfer function is of the form
the move suppression coefficient can be made indepen-
dent of the process gain (Lalonde and Cooper, 1989;                              y(z)    D(z)G(z)     D̃(z)G̃(z)
McIntosh et al., 1991). Gain-scaling is a term coined                                 )            )                     (28)
                                                                                ysp(z) 1 + D(z)G(z) 1 + D̃(z)G̃(z)
to represent a transformation where a mathematical
expression is stripped of the effects of process gain for
analysis independent of gain. For example, gain-scaling              By casting the move suppression, λ, as a scaled
of the move suppression coefficient can be done by
expressing it as a product of a scaled move suppression           coefficient, f, times the square of the process gain, K2p,
                                                                  the SISO DMC response looks similar for all first-order
coefficient, f, and the square of the process gain, K2p.
Note that this approach is restricted to DMC applied              systems when the time dimension is scaled appropri-
to SISO systems and cannot be directly extended to                ately. In other words, eq 28 shows that, by gain-scaling
multivariable systems.                                            λ as in eq 22, the closed-loop performance is rendered
  Consider a form of the move suppression coefficient             independent of the process gain. As a result, derivation
given by                                                          of an analytical expression for λ yields a scaled coef-
                                                                  ficient, f, that is a function of parameters other than
                          λ ) fK2p                        (22)    the process gain.

The step response coefficient of any linear system can            Derivation of an Analytical Expression for λ
be written as
                                                                    The analysis of the gain-scaled system matrix, ÃTÃ,
                         ai ) Kpãi                       (23)    involves the development of an approximate form of the
                                                                  gain-scaled system matrix. Such an approximation is
where ãi represents the part of the unit step response           made possible with the use of a FOPDT model ap-
coefficient that is independent of the process gain, Kp.          proximation of the process. It must be emphasized that
Using eq 22 and eq 23, it is possible to separate the gain-
                                                                  the use of this FOPDT approximation is employed only
related effects from the first row elements, ci, of the
                                                                  in the derivation of the analytical expression for λ. The
pseudoinverse matrix:
                                                                  examples presented later in this work all use the
                                                                  traditional DMC step response matrix of the actual
ci ) ith first row element of {(ATA + λI)-1AT}
                                                                  process upon implementation.
  ) ith first row element of {(K2pÃTÃ + fK2pI)-1KpÃT}            An Approximate Form of the ÃTÃ Matrix. A
  ) 1/Kp × ith first row element of {(ÃTÃ + fI)-1ÃT}           FOPDT model with zero-order hold is represented by a
                                                                  discrete transfer function as
  ) c̃i/Kp                                         (24)

Here, Ã is the gain-scaled dynamic matrix, ÃTÃ is the                                        Kp(1 - e-(T/τp))z-k
                                                                                    H0Gp(z) )                            (29)
gain-scaled system matrix, and c̃i is the ith first row                                          (1 - e-(T/τp)z-1)
element of the gain-scaled pseudo-inverse matrix.
  Substituting eq 23 and eq 24 in eq 21 shows that the
gain dependence of the SISO DMC transfer function is              where Kp is the process gain, τp is the overall process
separable from the dependence on remaining process                time constant, T is the discrete sample time, and k is
and controller parameters:                                        the effective discrete dead time given by

                ( ( ))
D(z) ) u(z)/e(z) )                                                                          k ) θp/T + 1                 (30)
                          1
                                                     ) D̃(z)/Kp
                               P                                  In eq 30, θp is the effective dead time in the process.

                1       N-1
                            ∑   c̃iãi+j                          From eq 29, the step response coefficients of a FOPDT
                      + ∑
                            i)1                                   process are given by
   Kp(1 - z )
           -1
                                         - ãj z-j
               P               P

              ∑               ∑
                                                                            {
                        j)1
                  c̃i              c̃i                                      0                            0ejek-1
              i)1             i)1                                   ãi )                                                (31)
                                                                            (1 - e-(T/τp)(j-k+1))        kej
                                                          (25)

Similarly, the gain dependence of a linear process                  Using a FOPDT model approximation of the process
transfer function is separable from the remaining                 and the gain-scaled step response coefficients from eq
process parameters:                                               31, the dynamic matrix from eq 9 has the form
736 Ind. Eng. Chem. Res., Vol. 36, No. 3, 1997

                 e

                 e                   e

                 e                   e                          e

             e                   e                      e                                e

         e                       e                      e                            e

For this form of the dynamic matrix, the gain-scaled
system matrix, ÃTÃ, in the DMC control law has the
form

                         e                     e            e                    e       e
                                                                                                          Figure 5. Comparison of the true and approximate condition
                                                                                                          numbers of the ÃTÃ matrix for large values of the prediction
                     e       e                      e                            e       e                horizon, P.

                                                                                                          The validity of the approximation in eq 36 is explored
                     e       e                 e            e                        e
                                                                                                          later in this section.
                                                                                                            The other terms of the ÃTÃ matrix can be ap-
                                                                                                          proximated using a similar procedure. The final ap-

                                                                                                          [                                                                       ]
                                                                                                          proximate form of the matrix that results is
    An approximate form of the gain-scaled system matrix
can be obtained by approximating individual terms of                                                      ÃTÃ =
the matrix in eq 33 for large values of the prediction                                                              3T           3T 3         3T
horizon, P, and small sample times, T (small T/τp). Let                                                    P-k-         +2 P-k-      + P-k-
                                                                                                                    2 τp         2 τp 2       2 τp
                                                                                                                                                  +1                  · · ·
R̃ij (i,j ) 1, 2, ..., M) be the term in the ith row and jth
column of the gain-scaled system matrix. The ap-                                                                    3T 3         3T           3T 1
                                                                                                              P-k-      + P-k-       +1 P-k-
                                                                                                                    2 τp 2       2 τp         2 τp 2
                                                                                                                                                  +                   · · ·
proximation of one such term, R̃11, is shown in eq 34.
    Recognizing that the summation terms in R̃11 are in                                                             3T           3T 1         3T
                                                                                                              P-k-      +1 P-k-      + P-k-
geometric progression results in the exact expression                                                               2 τp         2 τp 2       2 τp                    · · ·
                                                                                                                  ·            ·            ·                         ·
         P-k+1                                                                                                                                                                    M×M

                 ∑
                                                                                                                  ·            ·            ·                             ·
                         (1 - e-i(T/τp))2                                                                                                                                         (37)
                                                                                                                  ·            ·            ·                                 ·
R̃11 )

                                                                                                                           [                                      ]
                 i)1
         P-k+1                                                                                            Let â ) R̃11 = P - k - (3/2)(T/τp) + 2, then
    )            ∑
                 i)1
                         (1 - 2e         -i(T/τp)
                                                    +e          -2i(T/τp)
                                                                            )
                                                                                                                                             1
                                                                                                                           â            â-       â-1
                                                                                                                                             2
                                                                                                                                                          · · ·
                                               -(T/τp)                    -(P-k+1)(T/τp)
                                           2e                   (1 - e                         )                                    1                 3
   ) (P - k + 1) -                                                         +                                                   â-       â-1      â-
                                                    (1 - e-(T/τp))                                               ÃTÃ )            2                 2
                                                                                                                                                          · · ·
                                                                                                                                                                                  (38)
                                                                                                                                           3
                                                e-(2T/τp)(1 - e-2(P-k+1)(T/τp))                                            â-1          â-       â-2
                                                                                                                                           2
                                                                                                                                                          · · ·
                                                                                -(2T/τp)
                                                                                                   (34)
                                                                    (1 - e                 )                                ·            ·        ·       ·
                                                                                                                            ·            ·        ·         ·     M×M
                                                                                                                            ·            ·        ·           ·
With a FOPDT model approximation available, the
                                                                                                          Note that the approximate ÃTÃ matrix has a Hankel
prediction horizon, P, can be computed as the open-loop
                                                                                                          matrix form with the added feature that the elements
process settling time in samples as P ) (5τp/T) + k. This
                                                                                                          of every row successively decrease by 0.5 from left to
supports the findings of past researchers (Garcı́a and
                                                                                                          right. The approximate matrix of eq 38 is singular when
Morshedi, 1986; Maurath et al., 1988a; Lundström et
                                                                                                          M g 3. This supports the observation made by prior
al., 1995) that P should be large enough to include the
                                                                                                          researchers (Marchetti et al., 1983) that the ATA matrix
steady state effect of all past input moves.
                                                                                                          (hence, the ÃTÃ matrix) becomes increasingly singular
  For large values of prediction horizon, the term in eq
                                                                                                          for large values of the prediction horizon, P, and control
34 simplifies to
                                                                                                          horizon, M.
                                                                                                             Figure 5 provides evidence that the approximate gain-
                                               2e-(T/τp)                             e-(2T/τp)            scaled system matrix is a good approximation of the true
R̃11 = (P - k + 1) -                                        -(T/τp)
                                                                          +                        (35)
                                            (1 - e                    )         (1 - e-(2T/τp))           system matrix, especially for the analysis of the ÃTÃ +
                                                                                                          fI matrix in the DMC control law. Specifically, Figure
Notice that the approximation in eq 35 becomes increas-                                                   5 shows the condition number for the exact and ap-
ingly accurate as P increases and is exactly true for an                                                  proximate ÃTÃ + fI matrix as a function of the scaled
infinite horizon implementation of DMC. A first-order                                                     move suppression coefficient, f, for different choices of
Taylor series expansion, e-(T/τp) ) 1 - T/τp, that is valid                                               the control horizon. The prediction horizon employed
for small sample times (T/τp e 0.1), can be applied to                                                    in Figure 5 is (5τp)/T + k, which is the open-loop settling
eq 35 to yield                                                                                            time of the process based on a FOPDT model ap-
                                                                    3T
                                 R̃11 = P - k -                          +2                        (36)
                                                                    2 τp
Ind. Eng. Chem. Res., Vol. 36, No. 3, 1997 737
proximation of the process. This value of prediction                                 eq 45 by solving the system of equations that result from
horizon is recommended in the proposed tuning strategy                               the top left partitioned block using eq 38 and eq 42.
based on closed-loop stability considerations (Garcı́a and                           Thus, a and b are given by
Morari, 1982; Clarke and Scattolini, 1991; Rawlings and
Muske, 1993; Scokaert and Clarke, 1994). From Figure
5 it can be seen that, for large values of prediction
horizon, the condition number of the approximate
                                                                                                         a)M â-     {          (M - 1)
                                                                                                                                  2         }
system matrix in eq 38 closely follows the true condition                                                          Mx(M - 1)(M + 1)
number.                                                                                                b)-                                                    (46)
   QR Factorization of the Approximate ÃTÃ + fI                                                                              2x12
Matrix. Since the approximate form of the system
matrix (eq 38) was shown above to be a reasonable                                      Now, the ÃTÃ + fI matrix, to be inverted in DMC

                                                                                          [                                                             ]
approximation of the true system matrix (eq 33), the                                 control law, has the form
following treatment of the ÃTÃ + fI matrix is based on
this approximate form.                                                               ÃTÃ + fI )
   Consider two linearly independent M-vectors:                                                                1
                                                                                              â+f      â-                      â-1
                                                                                                               2
                                                                                                                                                · · ·
                            T
                      h
                      h 1 ) (1 1 1 ‚‚‚ 1)1×M                                                       1                                3
                                                                                              â-       â-1+f                   â-
                                                                                                   2                                2
                                                                                                                                                · · ·
                                                                                                                                                              (47)
                      T                                                                                        3
                h
                h 2 ) (0 1 2 ‚‚‚ M - 1)1×M                                   (39)             â-1      â-                      â-2+f
                                                                                                               2
                                                                                                                                                · · ·
The approximate ÃTÃ matrix (eq 38) can be written in                                         ·
                                                                                               ·
                                                                                                               ·
                                                                                                               ·
                                                                                                                                 ·
                                                                                                                                 ·
                                                                                                                                                ·
                                                                                                                                                  ·     M×M
terms of these vectors as                                                                      ·               ·                 ·                  ·
                                                                                     Adopting the factored form in eq 45 and eq 46, eq 47 is
                                       h 1T + h
                            ÃTÃ ) νj‚h      h 1‚νjT                        (40)    written as
where

                                                                                      [                                    ]
                                                                                     ÃTÃ + fI ) [ej 1 ej 2 ... 0]M×M ×
                                      â    1
                                  νj ) hh - h                                (41)      a+f         b
                                                                                                         ·
                                                                                                                   0
                                      2 1 2 2
                                             h                                                           ·
                                                                                       b           f
                                                                                                         ·
Hence, h        h 2 form a basis for the approximate ÃTÃ
        h 1 and h                                                                                                                [ej 1 ej 2 ... 0]TM×M (48)
                                                                                       · · ·           · · ·       · · ·
matrix.
                                                                                       0                           fI
                                                                                                         ·
  A Gram-Schmidt orthonormalization of h1 and h2                                                         ·                 M×M
                                                                                                         ·
yields the orthonormal basis for ÃTÃ:
                                                                                     Equation 48 can now be used to determine explicit
                       T      1                                                      analytical expressions for the eigenvalues of ÃTÃ + fI.
                   ej 1    )    (1 1 1 ‚‚‚ 1)1×M                                     From eq 47 it is clear that the approximate form of ÃTÃ
                             xM                                                      + fI has eigenvalues µ ) f with multiplicity (M - 2).
                                                                                     Therefore

          x
                   12
ej 2T )                       (0 1 2 ‚‚‚ M - 1)1×M -
              M(M - 1)(M + 1)                                                                                          µmin ) f                               (49)
                        (M - 1)
                                (1 1 1 ‚‚‚ 1)1×M (42)                                  The remaining two eigenvalues are found from the
                           2
                                                                                     top left partitioned block as the µ that satisfies
Therefore, h
           h 1 and h
                   h 2 can be QR factored as

                    h1 h
                   [h  h 2]M×2 ) [e1 ej 2]M×2R2×2                            (43)                      |ab + f - µ             b
                                                                                                                               f-µ
                                                                                                                                   )0   |                     (50)

where R is upper triangular and invertible. Using eq                                 A solution to eq 50 (using eq 46) yields the larger of the
43, eq 40 can be transformed to                                                      two eigenvalues as

           h 1T + h
ÃTÃ ) νj‚h      h 1‚νjT )                                                          µmax )

                           [ej 1 ej 2]M×2      [
                                               a
                                               b
                                                     b
                                                     0   ]
                                                         2×2
                                                                          T
                                                               [ej 1 ej 2]2×M (44)   Mâ+2f -
                                                                                               M(M-1)
                                                                                                 2
                                                                                                      +
                                                                                                               x    M2â2-M2(M-1)â+
                                                                                                                        2
                                                                                                                                  (M-1)M2(2M-1)
                                                                                                                                        6

where a and b are simple linear functions of â. Equa-                                                                                                         (51)
tion 44 can also be written as
                                                                                        Alternatively, the minimum and maximum eigenval-
  T                                                                                  ues of ÃTÃ + fI can be determined by reasoning. Note
à à ) [ej 1 ej 2 ... 0]M×M ×

   [                                       ]
                                                                                     that the ÃTÃ matrix (eq 38) is nearly singular for M )
                                                                                     2 and is perfectly singular for M > 2. Therefore, the
      a        b
                            ·
                            ·      0                                                 minimum absolute eigenvalue of ÃTÃ for M g 2 is close
                                                                                     to or exactly zero. When a constant quantity, f, is added
      b        0
                            ·
                                                   [ej 1 ej 2 ... 0]TM×M (45)        to the leading diagonal of such a matrix, all its eigen-
      · · ·               · · ·    · · ·                                             values shift by that quantity (Hoerl and Kennard, 1970;
      0                            0                                                 Ogunnaike, 1986). Hence, the minimum absolute eigen-
                            ·
                            ·              M×M
                            ·                                                        value of the resultant ÃTÃ + fI matrix, µmin, is equal to
Analytical expressions for a and b can be obtained from                              f (eq 49).
738 Ind. Eng. Chem. Res., Vol. 36, No. 3, 1997

   Analytical expressions for the maximum eigenvalue                          (M - 1)(2M - 1)
can be derived for the square matrix, ÃTÃ + fI, with        â2 - (M - 1)â +
                                                                                     6
                                                                                              =

                                                                                                    (           )
successively increasing dimensions (M × M). By rec-                                       2                      2
                                                                                   (M - 1)       (M - 1)
ognizing that the various coefficients in the analytical           â2 - (M - 1)â +          ) â-                     (55)
expressions follow a series pattern that is a function of                             4             2
the matrix dimension (M × M), a general formula for
the maximum eigenvalue can be obtained as a function          With this simplification, the expression for the gain
of M. This general analytical expression is identical to      scaled coefficient becomes

                                                                                         (              )
the analytical expression obtained in eq 51.
                                                                                     M    (M - 1)
   Formulation of the Analytical Expression for λ.                              f)     â-                            (56)
Past researchers (e.g., Ogunnaike, 1986) have indicated                              c       2
that the move suppression coefficient, λ, serves a dual
purpose in the DMC control law. Its primary role in              Past researchers (Maurath et al., 1988a,b; Callaghan
                                                              and Lee, 1988; Farrell and Polli, 1990) have indicated
DMC is to suppress otherwise aggressive controller
                                                              typical condition numbers for a moderately ill-condi-
action when M > 1. Additionally, λ serves to improve
                                                              tioned DMC system matrix. Condition numbers re-
the conditioning of the system matrix by rendering it
                                                              ported range from about 100 for single-input single-
more positive definite.                                       output systems to about 2000 for multivariable systems.
   One premise underlying this work is that both these        Apart from the variety of systems considered, the
effects are interrelated. When λ is increased, the            approximate relation, c ∝ M(P - k)/f, indicates that
manipulated input move sizes decrease, as does the            some of the differences are due to the different predic-
condition number. Clearly, if the effect of λ on the          tion and control horizons used by the individual design-
condition number of the system matrix can be analyti-         ers. In this work, a condition number of 500 is selected
cally expressed, then this condition number can be            to represent the upper allowable limit of ill-conditioning
maintained within specified bounds by an appropriate          in the system matrix (corresponding to modest control
choice of λ. An upper bound on the condition number           effort).
would, in turn, prevent the manipulated input move               The choice of the condition number, and hence the
sizes from becoming excessive.                                upper allowable limit of ill-conditioning in the system
   The condition number is defined for a square matrix        matrix, lies with the individual designer. Depending
                                                              on the criterion for good closed-loop performance for a
as
                                                              given application, the condition number can be conve-
                                                              niently fixed. The choice of a condition number of 500
                            |µmax|                            was motivated by the rule of thumb that the manipu-
                       c)                              (52)
                            |µmin|                            lated input move sizes for a change in set point should
                                                              not exceed 2-3 times the final change in manipulated
                                                              input (Maurath et al., 1985; Callaghan and Lee, 1988).
where µmax and µmin represent the maximum and
                                                              However, if a faster or slower closed-loop response is
minimum eigenvalues of the matrix. From eqs 49, 51,
                                                              more desirable, a larger or smaller condition number
and 52, the condition number for the ÃTÃ + fI matrix        respectively, can be used instead.
is                                                               Substituting the expression for â in eq 56, with P )

      (
                                                              (5τp)/T + k, the analytical expression for the move
     1            M(M - 1)                                    suppression coefficient, λ, is given by
c)      Mâ + 2f -

                                                                                     (                      )
     2f              2
                           +

                                                   )                             M 7 τp     (M - 1)
          M
           x   2
              â - (M - 1)â +
                             (M - 1)(2M - 1)
                                    6
                                             (53)
                                                                           f)
                                                                                500 2 T
                                                                                        +2-
                                                                                               2

                                                                                         λ ) fK2p                    (57)
An interesting approximate relation from eq 53 is that
the condition number of the DMC gain-scaled system            The analytical expression for λ in eq 57, applied in
matrix, c ∝ M(P - k)/f.                                       conjunction with the recommended values for the other
  Equation 53 is rearranged to give an expression for         tuning parameters, gives the tuning strategy for DMC
the scaled move suppression coefficient as                    with M > 1.
                                                                Note that eq 57, which computes the move suppres-

f)
     M
      (
     2c
        â-
           (M - 1)
              2
                   +
                                                              sion coefficient, does not contain a dead time term. The
                                                              reason is not intuitively obvious but is an interesting

                                                   )
                                                              one. The elements, and the condition number, of the

           x                     (M - 1)(2M - 1)              (ATA + λI) matrix depend on the number of non-zero
              â2 - (M - 1)â +                    (54)         terms in the dynamic matrix, A. This is clearer from
                                        6
                                                              eq 33, where the first term of the (ATA + λI) matrix is
                                                              a series summation performed over the P - k + 1 non-
where â ) P - k - (3/2)(T/τp) + 2. Equation 54
                                                              zero terms in the dynamic matrix A. (Equation 37 also
expresses the gain-scaled coefficient as a function of the    conveys the dependence of condition number of (ATA +
condition number of ÃTÃ + fI, the control horizon, and      λI) on P - k, rather than P alone). Additionally, the
FOPDT model parameters other than the process gain.           choice of prediction horizon as P ) (5τp/T) + k causes
  Evaluation of f in eq 54 can be simplified by recogniz-     the elements of (ATA + λI) to depend on P - k ) (5τp/
ing the contribution of each term to the value of f. The      T). Hence, the condition number of (ATA + λI) and the
last term within the square root in eq 54 can be modified     move suppression from eq 57, are dependent on (5τp/T)
to ease evaluation:                                           and are independent of dead time. Of course, this is
Ind. Eng. Chem. Res., Vol. 36, No. 3, 1997 739

                                                                         [                                       ]
true only if the prediction horizon is chosen as P ) (5τp/       controlled variable weights, ΓTΓ, has the form
T) + k (settling time of the process).

                                                                             γ21IP×P       0             0
                                                                                       ·             ·
Extension of Tuning Strategy to Multivariable                                          ·             ·
Systems                                                                      · · ·
                                                                                       ·
                                                                                       ·   · · ·
                                                                                                     ·
                                                                                                     ·   · · ·
                                                                   T
   The selection of tuning parameters for multiple-input                                   γ22IP×P
                                                                                       ·             ·
                                                                  Γ Γ) 0               ·             ·   0                 (61)
multiple-output (MIMO) DMC is significantly more                                       ·             ·
challenging for the practitioner. The analytical expres-                     · · ·     ·   · · ·     ·   · · ·
sion that computes the move suppression coefficient (eq
                                                                                       ·             ·
                                                                             0         ·   0         ·
                                                                                                         ·
                                                                                                           ·     P‚R×P‚R
57), developed in the previous section for SISO DMC,                                   ·             ·       ·
provides the foundation upon which a similar analytical
expression can be developed for MIMO DMC. This is                Hence, the controlled variable weights are γ2i (i ) 1, 2,
possible in a straightforward fashion even though the            ..., R).
performance objective for MIMO DMC is defined over                   Similar to SISO DMC, adjustable parameters that can
several manipulated inputs and measured outputs and              be used to alter closed-loop performance for MIMO DMC
results in a more complex MIMO DMC control law.                  include the prediction horizon, P, the control horizon,
   For a system with S manipulated inputs and R                  M, the model horizon, N, the sample time, T, and the
measured outputs, the MIMO DMC performance objec-                move suppression coefficients, λ2i . In addition, MIMO
tive (cost function) has the form (Garcı́a and Morshedi,         DMC has yet another set of adjustable parameters in
1986)                                                            the form of the controlled variable weights, γ2i .
                                                                     Just as with the SISO case, practical limitations often
                j ]TΓTΓ[ej - A∆u
Min J ) [ej - A∆u                      j ]TΛTΛ[∆u
                               j ] + [∆u         j]              restrict the availability of sample time, T, as a tuning
∆u
 j                                                               parameter for MIMO DMC (Franklin and Powell, 1980;
                                               (58)
                                                                 A° ström and Wittenmark, 1984). The model horizon, N,
                                                                 is also not an appropriate tuning parameter since
In eq 58, ej is the vector of predicted errors for the R         truncation of the model horizon can lead to unpredict-
measured outputs over the next P sampling instants,              able closed-loop performance (Lundström et al., 1995).
A is the multivariable dynamic matrix, and ∆u    j is the        As demonstrated earlier for SISO DMC, the control
vector of M moves to be determined for each of the S             horizon, M, does not have a significant effect on closed-
manipulated inputs. ΛTΛ is a square diagonal matrix              loop performance, especially in the presence of move
of move suppression coefficients with dimensions (M‚S            suppression.
× M‚S). Similarly, ΓTΓ is a square diagonal matrix that
                                                                     The controlled variable weights, γ2i , serve a dual
contains the controlled variable weights (equal concern
                                                                 purpose in MIMO DMC. These weights can be ap-
factors) with dimensions (P‚R × P‚R).
                                                                 propriately chosen by the user to scale measurements
   A closed-form solution to the MIMO DMC perfor-
                                                                 of the R measured outputs to comparable units. Also,
mance objective (eq 58) results in the MIMO DMC
                                                                 it is possible to achieve tighter control of a particular
control law (Garcı́a and Morshedi, 1986):
                                                                 measured output by selectively increasing the relative
                                                                 weight of the corresponding least-square residual. Hence,
                 j ) (ATΓTΓA + ΛTΛ)-1ATΓTΓej
                ∆u                                        (59)   controlled variable weights are usually specified by the
The MIMO DMC control law (eq 59) is similar to the               user for a certain application and cannot be employed
SISO DMC control law (eq 14), except for the form of             as the primary tuning parameters for MIMO DMC.
the move suppression coefficients, ΛTΛ, and the intro-               Analogous to SISO DMC, the move suppression
duction of controlled variable weights, ΓTΓ.                     coefficients can be conveniently used to fine tune MIMO
   The move suppression coefficients in multivariable            DMC for desired closed-loop performance. Since the
DMC follow a notation slightly different from SISO               dual benefit of the move suppression coefficients, λ2i , is
DMC. Λ is a square diagonal matrix with dimensions               again to improve the conditioning of the MIMO DMC
(M‚S × M‚S). This matrix can be divided into S2 square           system matrix (ATΓTΓA) and move size suppression for
blocks, each with dimensions (M × M). The leading                the S inputs, a strategy analogous to SISO DMC can
diagonal elements of the first (M × M) matrix block              be used to extend eq 57 to compute the move suppres-
along the diagonal of Λ are λ1, the leading diagonal             sion coefficients for MIMO DMC.
elements of the next (M × M) matrix block along the                  Building upon the analogy, an approximation of the
diagonal of Λ are λ2, and so on. All off-diagonal                MIMO DMC system matrix, (ATΓTΓA), has the form of
elements of the matrix Λ are zero. Hence, the matrix             a mosaic Hankel matrix (not shown here) comprised of

        [                                       ]
of move suppression coefficients, ΛTΛ, has the form              S2 Hankel matrix blocks. The S2 Hankel matrix blocks,
                                                                 each with dimensions (M × M), have a form identical
                                                                 to that obtained earlier (eq 47) from a similar ap-
         λ21IM×M          0             0                        proximation of the scaled SISO DMC system matrix,
                      ·             ·
                      ·             ·
                      ·             ·                            ÃTÃ.
                                                                     The impact of a change in the ith manipulated input
            · · ·     ·   · · ·     ·   · · ·
 T
                          λ22IM×M
                      ·             ·
Λ Λ) 0                ·             ·   0                 (60)   on all measured outputs is reflected in the ith diagonal
                      ·             ·                            Hankel matrix block. Hence, it is possible to select the
                                                                 ith move suppression coefficient, λ2i , such that the
            · · ·     ·   · · ·     ·   · · ·
                      ·             ·
            0         ·   0         ·
                                        ·
                                          ·     M‚S×M‚S          condition number of the ith diagonal matrix block is
                                                                 always bounded by a fixed low value. By holding the
                      ·             ·       ·
Thus, in the MIMO DMC control law (eq 59), the move              condition number of the ith diagonal matrix block at a
suppression coefficients that are added to the leading           low value, desirable closed-loop performance is achieved
diagonal of the multivariable system matrix, (ATΓTΓA),           while preventing the ith manipulated input move size
are λ2i (i ) 1, 2, ..., S). Similarly, the matrix of             from becoming excessive.
740 Ind. Eng. Chem. Res., Vol. 36, No. 3, 1997

  With this understanding, an analytical expression                               An additional criterion for sample time selection is
that computes the move suppression coefficients for                            that the sampling rate should be high enough to sample
MIMO DMC can be obtained as                                                    two or three times per effective dead time of the process,
                                                                               T e 0.5θp (Seborg et al., 1986). This criterion is not a

                    [ (                                              )]
                                                                               stringent requirement for DMC. However, if sample
          M    R
                                            3 τpij          (M - 1)
λ2i   )       ∑
          500 j)1
                    γ2j Kp2ij   P - kij -
                                            2 T
                                                     +2-
                                                                 2
                                                                               time can be chosen, the prudent approach is to pick the
                                                                               largest T that satisfies both criteria mentioned above.
                                                     (i ) 1, 2, ..., S) (62)   Once T is fixed, the discrete dead time, k, is calculated
                                                                               from the effective dead time of the process, θp in step 3.
Using eq 62, the S move suppression coefficients, one                             Step 4 computes a model horizon, N, and a prediction
for each input, can be computed for a given sampling                           horizon, P, from τp, θp, and T. Note that both N and P
time, T, control horizon, M, and controlled variable                           cannot be selected independent of the sample time, T.
weights, γi.                                                                   For DMC, it is imperative that N be equal to the open-
  In eq 63, it is not possible to substitute P ) (5τp)/T +                     loop process settling time in samples to avoid truncation
k as was done for the SISO case since the MIMO DMC                             error in the model prediction (Lundström et al., 1995).
architecture requires a single value of the prediction                         For computing P, a rigorous treatment by past research-
horizon to be selected for all manipulated input and                           ers (Garcı́a and Morari, 1982; Clarke and Scattolini,
measured output pairs. Nevertheless, the analytical                            1991; Rawlings and Muske, 1993; Scokaert and Clarke,
expression for MIMO DMC is similar in form to that                             1994) has shown that a larger prediction horizon
obtained for SISO DMC (eq 57). In fact, for a single-                          improves nominal stability of the closed loop. For
input single-output process, the analytical expression                         practical applications this translates to using a reason-
for MIMO DMC (eq 62) reduces to the analytical                                 ably large but finite P. Bearing in mind this important
expression for SISO DMC (eq 57) when P ) (5τp)/T + k.                          requirement for stability, P is also set equal to the open-
                                                                               loop process settling time in samples. With a FOPDT
                                                                               model approximation available, prediction horizon and
Implementation of the DMC Tuning Strategy                                      model horizon can both be computed as P ) N ) (5τp/
                                                                               T) + k.
   The proposed DMC tuning strategy, which includes
the analytical expression for the move suppression                                In an industrial setting, there can be barriers to
coefficient, λ, is presented in Table 1. This tuning                           selecting different P and T values for different SISO
strategy can be applied to unconstrained DMC in closed                         DMC controllers. Also, the MIMO DMC architecture
loop with a broad class of SISO processes that are open-                       requires that a single P and a single T be selected for
loop stable, including those with challenging control                          all manipulated input and measured output pairs. If
characteristics such as high process order, large dead                         the user has the freedom to select a sample time in such
time, and noniminimum phase behavior.                                          a scenario, a conservative choice would be the smallest
                                                                               value of T that satisfies eq 63 for all individual ma-
   Step 1 (Table 1) involves the identification of a first
                                                                               nipulated input and measured output pairs. Based on
order plus dead time (FOPDT) model approximation of
                                                                               this sample time (or the fixed sample time), a single
the process. An accurate identification of the FOPDT
                                                                               prediction horizon can be computed as P ) max {(5τpij)/T
model parameters is essential to the success of this
                                                                               + kij (i ) 1, 2, ..., S; j ) 1, 2, ..., R)} for a system with S
tuning strategy. Since a model is only as good as the
data it fits, it is necessary that the input-output data                       manipulated inputs and R measured outputs.
used for model fitting be rich in dynamic information                             Step 5 requires the specification of a control horizon,
of the process. Typically, this is achieved by perturbing                      M, and the computation of an appropriate move sup-
the process to obtain a signal-to-noise ratio greater than                     pression coefficient, λ. Recommended values of M range
10:1 (Seborg et al., 1986). Additionally, the data must                        from 1 through 6. However, selecting M > 1 provides
be collected over a reasonable period of time to capture                       advance knowledge of the impending manipulated input
the complete process dynamics. An FOPDT fit thus                               moves by the controller and can be very useful to the
obtained often reasonably describes the process gain,                          practitioner (Ogunnaike, 1986). Also, Rawlings and
Kp, overall time constant, τp, and effective dead time,                        Muske (1993) have shown that the stability of infinite
θp, of higher-order processes.                                                 horizon MPC can be guaranteed if M is greater than or
   Step 2 involves the selection of an appropriate sample                      equal to the number of unstable modes in the process.
time, T. In most practical applications the user does                             The strategy for computing λ depends on the choice
not have the freedom to choose sample time. The tuning                         of M. As demonstrated earlier, with M ) 1 the need
strategy does require that T be known. In cases where                          for a move suppression is obviated and λ is set equal to
the designer can select the sample time, certain rules                         zero. However, if M > 1, a scaled move suppression
of thumb that guide in its selection are available. Rules                      coefficient, f, is first computed from an analytical
that select T for a desired closed-loop bandwidth, ωB,                         expression (eq 57). The product of the scaled move
have been proposed in the past, e.g., T e 2π/10ωB                              suppression coefficient, f, and the square of the process
(Middleton, 1991). Since the choice of T is related to                         gain, K2p, determines an appropriate value for λ.
the overall time constant, τp, and the effective dead time                        Step 6 summarizes the DMC tuning parameters. For
of the process, θp (Seborg et al., 1986), the estimated                        certain applications, more specific or stringent perfor-
FOPDT model parameters provide a convenient way to                             mance criteria regarding the manipulated input move
select T. A good rule of thumb is to select T such that                        sizes or the nature of the measured output response may
the sampling rate is not less than 10 times per time                           apply. For such cases, it may be necessary to fine tune
constant. Hence, for DMC, the recommended sample                               DMC for desired performance by altering P and λ from
time is                                                                        the starting values given by the tuning strategy. The
                                                                               recommended approach is to increase λ for smaller move
                                   T e 0.1τp                           (63)    sizes and slower output response.
You can also read