Robust Multivariable Control of Ill-Conditioned Plants – A Case Study for High-Purity Distillation

Distillation is one of the most important unit operations in the chemical industry. Among various distillation operations, control of high-purity column poses difficult control due to a number of characteristics of these systems, including strong directionality, ill-conditioning and strongly nonlinear behavior. At the same time, the potential benefits that can be obtained through tight and economic control of the product compositions are very large. This is due to different reasons including the large energy consumption required by the columns and the market requirements which are becoming stricter and stricter. Because of these obvious features of high-purity distillation, this type of column has been studied extensively. Control systems for chemical processes are typically designed using an approximate, linear, time-invariant model of the plant. The actual plant may differ from the nominal model due to many sources of uncertainty, such as nonlinearity, the selection of low-order models to represent a plant with inherently high-order dynamics, inaccurate identification of model parameters due to poor measurements or incomplete knowledge, and uncertainty in the manipulative variables. Considering the differences between the actual plant and nominal model, it is necessary to insure that the control system will be stable and meet some predetermined performance criteria when applied to the actual plant. The identification and control of distillation columns have been subjects of frequent study due to the illconditioned nature of the distillation process. An ill-conditioned plant is very close to singular, and unless care is taken, very small errors can make the model useless. In distillation, this means that a model may have features that are in conflict with physical knowledge (Luyben, 1987; Jacobsen & Skogestad, 1994; Böling & Häggblom, 1996). In addition, ill-conditioned dynamics of high-purity distillation columns leads to high sensitivity to uncertainties in the manipulated variables (Skogestad & Morari, 1988). This effect causes even small errors in the manipulated variables show significant deterioration of the product quality, a fact which explains why open-loop control of high-purity distillation columns is hardly ever satisfactory. The model of a high-purity distillation process has a steady-state gain matrix with a high condition number. The gain matrix is almost singular and its determinant may be affected by quite small model errors, and if


Introduction
Distillation is one of the most important unit operations in the chemical industry. Among various distillation operations, control of high-purity column poses difficult control due to a number of characteristics of these systems, including strong directionality, ill-conditioning and strongly nonlinear behavior. At the same time, the potential benefits that can be obtained through tight and economic control of the product compositions are very large. This is due to different reasons including the large energy consumption required by the columns and the market requirements which are becoming stricter and stricter. Because of these obvious features of high-purity distillation, this type of column has been studied extensively. Control systems for chemical processes are typically designed using an approximate, linear, time-invariant model of the plant. The actual plant may differ from the nominal model due to many sources of uncertainty, such as nonlinearity, the selection of low-order models to represent a plant with inherently high-order dynamics, inaccurate identification of model parameters due to poor measurements or incomplete knowledge, and uncertainty in the manipulative variables. Considering the differences between the actual plant and nominal model, it is necessary to insure that the control system will be stable and meet some predetermined performance criteria when applied to the actual plant. The identification and control of distillation columns have been subjects of frequent study due to the illconditioned nature of the distillation process. An ill-conditioned plant is very close to singular, and unless care is taken, very small errors can make the model useless. In distillation, this means that a model may have features that are in conflict with physical knowledge (Luyben, 1987;Jacobsen & Skogestad, 1994;Böling & Häggblom, 1996). In addition, ill-conditioned dynamics of high-purity distillation columns leads to high sensitivity to uncertainties in the manipulated variables . This effect causes even small errors in the manipulated variables show significant deterioration of the product quality, a fact which explains why open-loop control of high-purity distillation columns is hardly ever satisfactory. The model of a high-purity distillation process has a steady-state gain matrix with a high condition number. The gain matrix is almost singular and its determinant may be affected by quite small model errors, and if determinant of the gain matrix of the model and that of the plant have different signs, no controller with integral action exists that can stabilize both the model and plant (Grosdidier et al., 1985). Many control design techniques have been applied to the high-purity distillation columns (e.g. Georgiou et al., 1988;Sriniwas et al., 1995;Christen et al., 1997;Shin et al., 2000;Razzaghi & Shahraki, 2005Biswas et al., 2009). Some possible improvements for linear multivariable predictive control of high-purity distillation columns are proposed by Trentacapilli et al. (1997) and a simple way of inserting a local model that contains part of the process nonlinearity into the controller is described also. In addition, a reliable model of the column is generally considered as a prerequisite for the design of efficient two-product control by multivariable methods. Another important aspect of distillation control design is the choice of a good configuration. In fact, poor control performance can result from the improper choice of manipulated/controlled variable pairing (Hurowitz et al., 2003). Some authors have been considered control configuration selection (Shinskey, 1984;Skogestad and Morari, 1987a;Finco et al., 1989;Stichlmair, 1995;Heath et al., 2000;Hurowitz et al., 2003;Luyben, 2005;Hori & Skogestad, 2007;Razzaghi & Shahraki, 2009), but there is no general agreement among these authors in choosing the best control configuration, however, a complete review in this field is performed by . The main works for selection of manipulated/controlled variable pairings have focused upon using controllability measures, such as relative gain array (Bristol, 1966) and structured singular value (Doyle, 1982). The relative gain array (RGA) provides a steadystate measure of coupling in multivariable systems and can be used to evaluate the steadystate coupling of configurations. RGA is still the most commonly used tool for control structure selection for single-loop controllers. Shinskey (1984) used the relative gain array to choose configuration which is applied widely in industry. Several authors such as  and Kariwala et al. (2006) have demonstrated practical applications of the RGA that it depends on the plant model only, that it is scaling independent and that all possible configurations can be evaluated base on the a single matrix. The structured singular value (SSV) approach provides necessary and sufficient conditions for robust stability and performance for the situation in which uncertainty occurs simultaneously and independently in various parts of the overall control system (e.g. input and output uncertainty) but the perturbation matrix is still norm-bounded. One of the most difficult steps in analysing the robust stability and performance of any control system is the specification of an estimate of the uncertainty associated with the nominal process model. This is a critical step because an overestimation of the model inaccuracy will lead to extensively poor control performance and an underestimation may lead to instability (McDonald et al., 1988). Several papers discuss ways in which model inaccuracy can be described and methods that can be used for assessing robust stability. The most common multivariable approaches that use singular values (Doyle and Stein, 1981;Arkun et al., 1984) and structured singular values assume that the actual plant can be described by a normbounded perturbation matrix in the frequency domain. In chemical process control, nonlinearity is one of the most significant sources of model inaccuracy. We usually have some knowledge about the structure of model inaccuracy due to nonlinearity, however, and this knowledge should be exploited in our robustness studies. In formulating the SSV problem, use of physically-based uncertainty description is important. Simplified models that predict gain and time constant changes as the process is perturbed over the expected operating regime can be used to characterise the uncertainty (McDonald et al., 1988).
The objective in this chapter is to show that acceptable closed-loop performance can be achieved for an ill-conditioned high-purity distillation column by use of the structured singular value . The distillation column model used in this case study is a high-purity column, referred to as "column at operating point A" by . Table  1 summarizes the steady-state data of the model in detail. The following simplifying assumptions are also made for the column: (1) binary separation, (2) constant relative volatility, and (3) constant molar flows. To include the effect of neglected flow dynamics, we will add uncertainty when designing and analysing controller.

Column data
Relative volatility

Process description
A simple two time-constant dynamic model presented by  is chosen as the basis for the controller design. The model is derived assuming the flow and composition dynamics to be decoupled, and then the two separate models for the composition and flow dynamics are simply combined. The nominal model of the column is given by where n is the number of trays in the column (N T -1). Fig. 1 shows a schematic of a binary distillation column that uses reflux and vapor boilup as manipulated inputs for the control of top and bottom compositions, respectively. This is denoted as the LV-configuration (structure). This structure is commonly used in industry for one-point composition control.

www.intechopen.com
However, severe interactions often make two-point control difficult with this configuration. Although the closed-loop system may be extremely sensitive to input uncertainty when the LV-configuration is used, while it is shown that it is possible to obtain good control behavior (i.e. good performance) with the LV-configuration when model uncertainty and possible changes in the operating point are included . The simultaneous control of overhead and bottoms composition in a binary distillation column using reflux and steam flow as the manipulated variables often proves to be particularly difficult because of the coupling inherent in the process. The result of this coupling, which cause the two control loops to interact, leads to a deterioration in the control performance of both composition control loops compared to their performance if the objective were control of only one composition. Since high-purity distillation columns can be very sensitive to uncertainties in the manipulated variables, it is important for successful implementation that a controller guarantees its performance in the presence of uncertainties. This particular design task is frequently solved by modeling a multiplicative uncertainty for a nominal plant model and subsequently calculating the controller using -synthesis (Doyle, 1982).  Fig. 2 shows general control problem formulation, where G is the generalized plant and C is the generalized controller. The controller design problem is divided into the analysis and synthesis phases. The controller C is synthesized such that some measure, in fact a norm, of the transfer function from w to z is minimized, e.g. the H ∞ -norm. Then the controller design problem is to find a controller C (that generates a signal u considering the information from v to mitigate the effects of w on z) minimizing the closed-loop norm from w to z. For the analysis phase, the scheme in Fig. 2 is to be modified to group the generalized plant G and the resulting synthesized controller C in order to test the closed-loop performance achieved with C. To get meaningful controller synthesis problems, weights on the exogenous inputs w and outputs z are incorporated. The weighting matrices are usually frequency dependent and typically selected such that the weighted signals are of magnitude one, i.e. the norm from w to z should be less than one. Once the stabilizing controller C is synthesized, it rests to analyze the closed-loop performance that it provides. In this phase, the controller for the configuration in Fig. 2 is incorporated into the generalized plant G to form the system N, as it is shown in Fig. 3. The expression for N is given by

General control problem formulation
where F l (G, C) denotes the lower Linear Fractional Transformation (LFT) of G and C. In order to obtain a good design for C, a precise knowledge of the plant is required. The dynamics of interest are modeled but this model may be inaccurate and may not reflect the changes suffered by the plant with time. To deal with this problem, the concept of model uncertainty comes out. The plant G is assumed to be unknown but belonging to a class of models, P, built around a nominal model G o . The set of models P is characterized by a matrix Δ, which can be either a full matrix or a block diagonal matrix that includes all possible perturbations representing uncertainty to the system. The general control configuration in Fig  The block diagram in Fig. 4 is used to synthesize the controller C. To transform it for analysis, the lower loop around G is closed by the controller C and it is incorporated into the www.intechopen.com generalized plant G to form the system N as it is shown in Fig. 5. The same lower LFT is obtained as in Eq.
(3) where no uncertainty was considered. To evaluate the relation between w  [w 1 w 2 ] T and z  [z 1 z 2 ] T for a given controller C in the uncertain system, the upper loop around N is closed with the perturbation matrix . This results in the following upper LFT: To represent any control problem with uncertainty by the general control configuration in Fig. 4, it is necessary to represent each source of uncertainty by a single perturbation block Structured uncertainty representation considers the individual uncertainty present on each input channel and combines them into one large diagonal block. This representation avoids the norm-physical coupling at the input of the plant that appears with the full perturbation matrix  in an unstructured uncertainty description. Consequently, the resulting set of plants is not so large as with an unstructured uncertainty description and the resulting robustness analysis is not so conservative (Balas et al., 1993).

Robust performance and robust stability
For obtaining good set point tracking, it is obvious that some performance specifications must be satisfied in spite of unmeasured disturbances and model-plant mismatch, i.e. uncertainty. The performance specification should be satisfied for the worst-case combination of disturbances and model-plant mismatch (robust performance). In order to achieve robust performance, some specifications have to be satisfied.  Fig. 5, || F u (N, Δ)|| ∞  1 and || Δ || ∞  1. The structured singular value is used as a robust performance index. To use this index one must define performance using the H ∞ framework. The H ∞ -norm of a transfer function G(s) is the peak value of the maximum singular value over all frequencies Uncertainties are modeled by the perturbations and uncertainty weights included in G. These weights are chosen such that || Δ || ∞  1 generates the family of all possible plants to be considered (Fig. 4). Δ may contain both real and complex perturbations, but in this case study only complex perturbations are used. The performance is specified by weights in G which normalized 2 w and 2 z such that a closed-loop H ∞ -norm from 2 w to 2 z of less than one (for worst-case Δ) means that the control objectives are achieved. Fig. 6 is used for robustness analysis where N is a function of G and C, and P  (|| Δ P || ∞  1) is a fictitious "performance perturbation" connecting 2 z to 2 w . Provided that the closed-loop system is nominally stable, the condition for robust performance (RP) is where diag{, } . P    is computed frequency-by-frequency through upper and lower bounds. Here we only consider the upper bound which is derived by the computation of non-negative scaling matrices D l and D r defined within a set D that commutes with the structure : . A detailed discussion on the specification of such a set D of scaling matrices can be found in Packard and Doyle (1993).

Design procedure
The design procedure of a control system usually involves a mathematical model of the dynamic process, the plant model or nominal model. Consequently, many aspects of the real plant behavior cannot be captured in an accurate way with the plant model leading to uncertainties. Such plant-model mismatching should be characterized by means of disturbances signals and/or plant parameter variations, often characterized by probabilistic models, or unmodelled dynamics, commonly characterized in the frequency domain. The modern approach to characterizing closed-loop performance objectives is to measure the size of certain closed-loop transfer function matrices using various matrix norms. Matrix norms provide a measure of how large output signals can get for certain classes of input signals. Optimizing these types of performance objectives, over the set of stabilizing controllers is the main thrust of recent optimal control theory, such as L 1 , H 2 , H ∞ and optimal control (Balas et al., 1993). Usually, high performance specifications are given in terms of the plant model. For this reason, model uncertainties characterization should be incorporated to the design procedure in order to provide a reliable control system capable to deal with the real process and to assure the fulfillment of the performance requirements. The term robustness is used to denote the ability of a control system to cope with the uncertain scheme. It is well known that there is an intrinsic conflict between performance and robustness in the standard feedback framework (Doyle and Stein, 1981;Chen, 1995). The system response to commands is an open-loop property while robustness properties are associated with the feedback. Therefore, one must make a tradeoff between achievable performance and robustness. In this way, a high performance controller designed for a nominal model may have very little robustness against the model uncertainties and the external disturbances. For this reason, worst-case robust control design techniques such as -synthesis, have gained popularity in the last thirty years.

Modeling of the uncertain system
Analyzing the effect of uncertain models on achievable closed-loop performance and designing controller to provide optimal worst-case performance in the face of the plant uncertainty are the main features that must be considered in robust control of an uncertain system.  recommended a general guideline for modeling of uncertain systems. According to this, three types of uncertainty can be identified: 1. Uncertainty of the manipulated variables which is referred to input uncertainty. 2. Uncertainty because of the process nonlinearity, and 3. Unmodelled high-frequency dynamics and uncertainty of the measured variables which is referred to output uncertainty. Fig. 7(a) shows a block diagram of a distillation column with related inputs (u, d) and outputs (y, y m ). In Fig. 7(b), we have added two additional blocks to Fig. 7(a). One is the controller C, which computes the appropriate input u based on the information about the process y m . The other block, Δ, represents the model uncertainty. Ĝ and G are models only, and the actual plant is different depending on Δ. Based on the measurements y m , the objective of the controller C is to generate inputs u that keep the outputs y as close as possible to their set points in spite of disturbances d and model uncertainty Δ. The controller C is often non-square, as there are usually more measurements than manipulated variables. For the design of the controller C, information about the expected model uncertainty should be taken into account. Usually, there are two main ways for adding uncertainty to a constructed model: additive and multiplicative uncertainty. Fig. 7(c) represents additive uncertainty. In this case, the perturbed plant gain G p will be G + Δ where Δ is unstructured uncertainty. Fig. 7(d) represents multiplicative uncertainty where the perturbed plant is equal to G (I + Δ). Here we will consider only input and output uncertainties: Input uncertainty-Input uncertainty always occurs in practice and generally limits the achievable closed-loop performance . Ill-conditioned plants can be very sensitive to errors in the manipulated variables. The bounds for the relative errors of the column inputs u are modeled in the frequency domain by a multiplicative uncertainty with two frequency-dependent error bounds u w . These two bounds are combined in the diagonal matrix uu Ww  I . In this case The value of the bound u W is almost very small for low frequencies (we know the model very well there) and increases substantially as we go to high frequencies where parasitic parameters come into play and unmodelled structural flexibility is common. If all flow measurements are carefully calibrated, an error bound of 10% for the low frequency range is reasonable (Christen et al., 1997). This error bound is not common among the researchers (e.g. , used an error bound of 20% at steady state). Higher errors must be assumed in the higher frequency range. Because of uncertain or neglected high-frequency dynamics or time delays, the input error exceeds 100%. The following weight is used as input uncertainty weight www.intechopen.com 11 0 () 0 . 1 1 u s ws s The weight is shown graphically as a function of frequency in Fig. 8.

Fig. 8. Input uncertainty weight w u ( jω) as a function of frequency.
Output uncertainty-Due to the nonlinear vapor/liquid equilibrium, the gains of the individual transfer functions between the two manipulated inputs and controlled outputs may change in opposite directions (gain directionality). This behavior can be described with independent multiplicative uncertainties for the two outputs of the model and a diagonal weighting matrix yy Ww  I . In mathematical form we can write For the low-frequency range, an uncertainty of 10% is assumed for the description of uncertainties in the measured outputs. The uncertainty weight is 1 180 () 0 . 1 12 . 5 y s ws s which has large gains in the high-frequency range that takes the effect of unmodelled dynamics into account. Performance-The performance weight used in this study is the same in . The weight is defined as 11 0 () 0 . 5 10 P s ws s 3.1 Controller  proposed two different approaches to tune controllers. The first approach is to fix the performance specification and minimize RP by adjusting the www.intechopen.com controller tunings. The performance requirement is satisfied if RP is less than one, and lower RP values represent a better design. The second approach is to fix the uncertainty and find what performance can be achieved. In this approach, we adjust the time constant in the performance weight to make the optimal RP values equal to one. The latter approach has two disadvantages: (1) it introduces an outer loop in the  calculations, and (2) it may be impossible to achieve RP equal to one by adjusting the time constant in the performance weight. Here the first approach is used for tuning the controller because of the mentioned disadvantages of the second approach. A diagonal PID controller based on internal model control (IMC) (Rivera et al., 1986) is used to investigate the process. Optimal setting for single-loop PID controller is found by minimizing RP . Furthermore, a -optimal controller is designed since it gives a good indication of the best possible performance of a linear controller.

Analysis of controller
Comparison of controller is based mainly on computing  for robust performance. The main advantage of using the μ-analysis is that it provides a well-defined basis for comparison. μanalysis is a worst-case analysis. It minimizes the H∞-norm with respect to the structured uncertainty matrix Δ. A worst-case analysis is particularly useful for ill-conditioned systems in the cross-over frequency range (Gjøsaeter and Foss, 1997). This is due to the fact that such systems may provide large difference between nominal and robust performance. The value of RP is indicative of the worst-case response. If RP > 1, then the "worst-case" does not satisfy our performance objective, and if RP < 1 then the "worst-case" is better than required by our performance objective. Similarly, if NP < 1 then the performance objective is satisfied for the nominal case. However, this may not mean very much if the system is sensitive to uncertainty and RP is significantly larger than one. It is shown that this is the case, for example, if an inverse-based controller is used for the distillation column . Controller was obtained by minimizing sup ω RP for the model using the input and output uncertainties and performance weight. The plots for RP for the -optimal controller are of particular interest since they indicate the best achievable performance for the plant. provides a much easier way of comparing and analyzing the effect of various combinations of controllers, uncertainty and disturbances than the traditional simulation approach. One of the main advantages with the -analysis as opposed to simulations is that one does not have to search for the worst-case, i.e. finds it automatically .

Synthesis of controller
The structured singular value provides a systematic way to test for both robust stability and robust performance with a given controller C. In addition to this analysis tool, the structured singular value can be used to synthesize the controller C. The robust performance condition implies robust stability, since Therefore, a controller designed to guarantee robust performance will also guarantee robust stability. Provided that the interconnection matrix N is a function of the controller C, theoptimal controller can be found by At the present time, there is no direct method to find the controller C by minimizing (16), however, combination of -analysis and H ∞ -synthesis which is called -synthesis or DKiteration (Zhou et al., 1996) is a special method that attempts to minimize the upper bound of . Thus, the objective function (16)  convergence in each step is assured, joint convergence is not guaranteed. However, DKiteration works well in most cases (Balas et al., 1993;Packard and Doyle, 1993). The optimal solutions in each step are of supreme importance to success with the DK-iteration. Moreover, when C is fixed, the fitting procedure plays an important role in the overall approach. Low order transfer function fits are preferable since the order of the H  problem in the following step is reduced yielding controllers of low order dimension. Nevertheless, the method is characterized by giving controllers of very high order that must be reduced applying model reduction techniques (Glover, 1984).

Simulation
Simulations are carried out with the nonlinear model of the column and using single-loop controller, which generally is insensitive to steady-state input errors . In addition, input and output uncertainties are included to get a realistic evaluation of the controller. Simulations are for both cases with and without uncertainty.

Model analysis 4.1 RGA-analysis of the model
Let  denote element-by-element multiplication. The RGA of the matrix G (Bristol, 1966) is defined as www.intechopen.com where g ij s are open-loop gain from the jth input to the ith output of the process. The RGA has been considered as an important MIMO system information for feedback control.
Controllers with large RGA elements should generally be avoided, because otherwise the closed-loop system is very sensitive to input uncertainty (Skogestad and Morari, 1987b). Fig.  9 shows the magnitude of the diagonal element of the RGA ( 11 ). As seen in the figure, the plant is ill-conditioned at low frequencies, while at higher frequencies, the value of the RGA-element drops. This says that only based on the RGA plot, making a decision on the ill-conditionedness of the control problem may be misleading. On the other hand, the bandwidth area is located in a frequency range where the RGA elements are small or at lower frequencies where the RGA elements are large.  as a function of frequency.

Ill-conditionedness and process gain directionality
The common definition of an ill-conditioned plant is that it has a model with a large condition number (  ). The condition number is defined as the ratio between the largest and smallest singular values ( /   ) of a process model. However, the condition number depends on the scaling of the process model. This problem arises from the scaling dependency of the Singular Value Decomposition (SVD). To eliminate the effect of scaling, the minimized condition number ( min  ) is defined as the smallest possible condition number that can be achieved by varying the scaling. Close relationship between min  and RGA is proposed by Grosdidier et al. (1985). For 2  2 systems 2 min 11 where the 1-norm of the RGA is defined as www.intechopen.com According to the above relationship, a 2  2 system with small RGA elements always has a small min  . In particular, if 11 01    the minimized condition number is always equal to one. A process model with a large span in the possible gain of the model is said to show high directionality and a process model with the smallest singular value equal to the largest singular value is said to show no directionality. Waller et al. (1994) suggest redefined definition of process directionality. The definition divides the concept of process directionality into two parts. The minimized condition number is connected to stability aspects, whereas the condition number of a process model scaled according to the weight of the variables is connected to performance aspects. Fig. 10 shows the largest and smallest singular values and condition number of the process model as a function of frequency. The condition number of the process is about 10 times lower at high frequencies than at low frequencies (steady state). Fig. 11(a) represents the values of  and min  as a function of frequency. Values of  and min  match each other from low to intermediate frequencies, but min  approaches one at high frequencies. For 2  2 systems (Grosdidier et al., 1985): Consequently, for 2  2 systems the difference between these quantities is at most one and 1  approaches min  as min  . Since 1  is much easy to compute than min  , it is the preferred quantity to use. In Fig. 11  approach one (after ω  20 rad/min). This is in agreement with the result obtained from 11 -vs-frequency plot (Fig. 9). Since min  is independent of scaling, therefore it is better to use min  instead of  , which is scale dependent. www.intechopen.com

Synthesis of the controller
The plots of the singular values of the sensitivity functions 1 () SG C   I demonstrate good disturbance rejection properties, which indicate the closed-loop system is insensitive to uncertainties in inputs and outputs ( Fig. 12(a)). The tracking properties of this controller are also adequate, which is illustrated by plots of the complementary sensitivity function, TS  I ( Fig. 12(b)). Up to the mid-frequency range, the singular values are close to one and the maximum of the upper singular values is slightly greater than one.  Table 2 summarizes the PID controller setting that is used for the column. Fig. 13 showsplots of the controller. From a maximum peak-value point of view, it is seen that both robust www.intechopen.com and nominal performance plots are less than one which satisfy the criterion. The plots approach 0.5 as frequency approaches infinity.  www.intechopen.com

Comparison with μ-optimal controller
Nominal and robust performance plots of the -optimal controller is shown in Fig. 14.
Comparison of nominal performance of the controllers shows that for the -optimal controller, the plot is nearly flat over a large frequency range which indicates that an optimal controller is achieved. Comparing robust performance of the controllers indicates that obtaining robust performance with the LV-configuration is also possible. This is also in agreement with the results presented by .

Simulations
Simulations of a set-point change in y D using the PID-and -optimal controllers are shown in Figs. 15 and 16, respectively. As it is seen, the introduced uncertainties do not seriously affect the performance of the -optimal controller, while for the PID-controller, the effect of uncertainties is more rather the -optimal controller. It should be noted that the reference signal is filtered by a prefilter with a time constant of 5 min. Fig. 16 also shows that the PID controller has a slow return to steady state. This is due to the high NP value at lower frequencies compared with the -optimal controller (Figs. 13 and 14). In Table 3, numerical values of for nominal and robust performance are presented.   Fig . 17 shows the closed-loop response of the -optimal controller to a 20% increase in feed flow rate. In Fig. 18, the closed-loop response for both controllers is shown simultaneously. As the figure shows, the PID controller needs considerably more times to reach steady state than the -optimal controller (see next page for the figures). Fig. 19 shows the effect of output uncertainty on closed-loop response of the -optimal controller. For the case that both input and output uncertainties are considered, the response is faster than for the case that only input uncertainty is considered, however, this difference is not so large. The reason for this again returns to the NP values at low frequencies.

Effect of output uncertainty
Thevalues of nominal performance for the case including both input and output uncertainties is close to the case where only input uncertainty included (Table 3) Fig. 17. Closed-loop response to a 20% increase in feed flow rate ( -optimal controller):  no uncertainty; ----10% uncertainty on inputs and outputs.

Discussion
The structured singular value ( ) is used to investigate the robust performance and robust stability of the PID controller. The control problem formulation used in this study is using weighted input and output uncertainties. Although other sources of uncertainty could be included, however, these two are the most severe uncertainties that may be considered. The inclusion of both input and output uncertainty prevents the control system from becoming sensitive to the uncertainties, as may happen with inverting controllers. The solution of the problem leads to the inequality of Eq. (8). The numerical solution of this design task is difficult. At present, there is no direct method to synthesize a -optimal controller, however, combination of -analysis and H  -synthesis which called -synthesis or DK-iteration, often yields good results. This algorithm has two drawbacks. Firstly, the algorithm cannot guarantee convergence, and secondly, the algorithm requires a scaling of the plant in each iteration step, which increases the order of the plant. The analysis advantageously avoids dealing explicitly with the bad condition of the plant.
With the -approach, the upper bound for the bandwidth of the control system is provided by the uncertainty model, whereas the lower bound is a matter of optimization. -synthesis is ideally suited to deal with complex uncertainty models which takes into account such aspects as various operating points. A difficulty that one may encountered in synthesis of controller is high computation time, because the approach requires scaling in each iteration. If, however, loop-shaping ideas are used to form the augmented plant, H ∞synthesis may be used to advantage. In this case, the results are as good as with thesynthesis, but are obtained with less numerical efforts (Christen et al., 1997). In this case study, the LV-configuration is used. The use of this configuration for columns with high condition number may be doubtful, but under special considerations, this configuration may yield acceptable performance. It is shown ) that it is possible to achieve good control behavior using the LV-configuration for twopoint composition control provided measurement delays are not too large (typically less than 1-2 min). In addition, severe interactions and poor control often reported with the LVconfiguration may be almost eliminated if the loops are tuned sufficiently tight. However, this does not imply that the LV-configuration is the best structure to use. Shinskey (1984) showed that the use of the (L/D)(V/B)-configuration is probably better in most cases, and in particular for columns with large reflux.

Concluding remarks
Based on a structured uncertainty model, which describes the column dynamics within the entire operating range, a decentralized PID controller is calculated using the -synthesis technique. The controller was found to be robust with respect to model-plant mismatch, provided the RGA values of the column transfer function are not too large in the cross-over frequency range. The response of the system is improved by using a -optimal controller. In spite of high condition number of the process, nominal and robust performance is achieved by insertion of input and output uncertainties in the control system and using the structured singular value to synthesis the controller. Good set-point tracking and disturbance rejection of the controller is observed by simulations that carried out for the closed-loop system. It was also shown that good control performance can be obtained by using the LV-configuration which is difficult to implement for two-point control. The obtained results also verify the findings of .