Investeşte în oameni! FONDUL SOCIAL EUROPEAN Programul Operaţional Sectorial pentru Dezvoltarea Resurselor Umane 2007 – 2013

Ing. Ionuţ MUNTEAN

PHD THESIS

MODELING, SIMULATION AND CONTROL OF DISTILLATION PROCESSES

SUMMARY

Thesis advisor

Prof. dr. ing. Mihail ABRUDEAN

TECHNICAL UNIVERSITY OF CLUJ-NAPOCA

FACULTY OF AUTOMATION AND COMPUTER SCIENCE

2011

Contents

1 Introduction 1

2 Separation processes 22.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.2 ACBT process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.3 Isotope enrichment unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

3 Modeling 43.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

3.2 Selecting a Suitable Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

3.3 Model Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3.4 Submodel-Library Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3.5 Model validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4 Full-order nonlinear observer 84.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4.2 State estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

5 Control of distillation processes 105.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

5.2 Decomposition of the control on two layers . . . . . . . . . . . . . . . . . . . . . . 10

5.3 Control of the ACBT process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

5.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

5.3.2 Stabilizing layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

5.3.3 Supervisory layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

5.3.4 Results and discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

6 Conclusions 13

References 15

Introduction

1 Introduction

Distillation is, after many years, still the most used method of performing separations in the chem-

ical and petrochemical industries, accounting for 95 % of all separations. On the other hand, its ther-

mal energy requirements are enormous, distillation processes being less than 10 % efficient [1]. This

has been translated into a constant interest from the academic environment to improve distillation,

where a search by “distillation” in Web of Science (WoS) returns over 60.000 articles and patents

with a rate of 200 new articles (on distillation and absorption) per year [2].

One way of saving energy in distillation is by using advanced control systems, which are able to

incorporate optimization algorithms in order to minimize the operational costs of the columns [1].

One advanced control system that has become an attractive feedback strategy, having found many

successful applications is model predictive control (MPC), also referred to as receding horizon control

[3]. In MPC schemes a process model, incorporated in an on-line open loop optimization problem, is

used to predict the effect of a finite number of future moves on the controlled variables. After having

determined the optimal control sequence for a given performance criterion, it then sends out only the

first change in each independent variable to be implemented, and repeats the calculation when the

next change is required.

Because of all the appealing features it provides, MPC has become a preferred control strategy

for a large number of processes. However, its main drawback lay in the time and expense required

for modeling and system identification, estimated to account for up to 80 % of the total design and

installation cost [1]. Reducing the modeling time is obtained with a model library, where previously

developed sub-models can be used to quickly agregate a model of a complex distillation process.

For this, a model library using first principle models (FP) has been created in gPROMS. FP models

are attractive because they are globally valid and therefore well suited for optimization. The models

developed in our work are aimed at being general control models, models capable of “fitting” most

of the distillation processes in use today, tailored in principle for industrial processes. In order to

prove the library’s general character, two processes have been considered to validate the models: a 3

distillation column process used to separate an echimolar azeotropic mixture of acetone, chloroform,

benzene and toluene (ACBT) and a laboratory scale 13C isotopic separation column.

To effectively implement MPC algorithms the knowledge of the mixture composition within the

column is required. Unfortunately, in practice, concentration measurements are seldom available

mainly because of the high price and complexity of the devices. The alternative is the use of a state

estimator and Kalman filters are the most used composition estimators in advanced control algorithms

for distillation columns. However, they have two main drawbacks: the need of model linearization

and the number of tuning parameters.

To overcome these problems, using the increased computational power available today, a full non-

linear observer proves a good and robust solution. The observer of Lang and Gilles [4] is approached

and a set of simulations have been ran in order to check its convergence.

1

Separation processes

Control has been approached from the plantwide perspective [5], where the control level is sep-

arated on two layers: the regulatory control and the supervisory control, where the MPC algorithm

was implemented. The starting point for plantwide control is to determine the degrees of freedom for

optimization. These are represented by the regulators from the regulatory layer whose setpoints will

be manipulated by the supervisory layer. Only then the objective function and the constraints from

the MPC can be properly defined. The implementation of the control strategy was done in DRTO

Toolbox [6] which has been extended, in order to be compatible with the modeling library, and then

tested on the ACBT process.

The thesis is structured into 6 chapters. After the introduction (chapter 1) follows, in chapter 2,

a detailed presentation of the two processes used as case studies throughout the thesis: the ACBT

process and the 13C isotopic separation column. In chapter 3 the equations and the assumptions

considered in developing the model library have been introduced, together with a literature review and

an overview of the most important concepts considered in the models. Chapter 4 presents a nonlinear

observer that has been selected in detriment to Kalman filters, based on its low number of design

parameters which can be tuned easily by simulation. In chapter 5 the plantwide control perspective is

being presented together with a complete control analysis of the ACBT process. Chapter 6 presents

the conclusions, personal contributions and states the future research directions.

2 Separation processes

2.1 Introduction

The chapter starts with a description of distillation and presents the role of each unit from a

distillation column. The main constructive techniques are also mentioned here stating the advantages

and disadvantages of each one. It then follows a rigorous presentation of two processes that have been

selected as case studies in the thesis: the ACBT process and the 13C isotope separation process. A

short overview of these two processes will be included next.

2.2 ACBT process

In order to separate an echimolar azeotropic mixture of acetone, chloroform, benzene and toluene

a three distillation column process with recycle has been selected.

The structure of the process has been determined by an optimization-based design procedure pre-

sented by Kraemer [7]. Here, the authors cover the whole design procedure of the process synthesis

framework, presented in the article, for a complete separation of an azeotropic four-component mix-

ture in a curved-boundary proces. By fully exploiting the special curvature of the distillation boundary

a complete separation was achieved in three columns. The recycle is required to shift the mass balance

line of the acetone column towards benzene/toluene. This three column structure provides, compared

2

2.3 Isotope enrichment unit

to other 4 column processes [8], a reduction in the investment cost and a reduction of the reboiler duty

close to 30 %. The final structure of the process is presented in figure 1.

Figure 1: Separation of the ACBT mixture [7]

2.3 Isotope enrichment unit

The column for the enrichment of 13C isotope studied in this paper, belongs to the National In-

stitute of Research and Development for Isotopes and Molecular Technologies (INCDTIM) Cluj-

Napoca, Romania. The column is being used as the first enrichment step for the 13C isotope from a

natural concentration in carbon monoxide of 1.1 % to a concentration above 3 %. Constructive details

were presented by Radoi et. al. [9] and a schematic representation of the plant is shown in Fig. 2.

The relation between the relative volatility (α) and the temperature for the carbon isotopes is given

by Axente [10]. To facilitate a reasonable separation factor the temperature should be maintained at a

value that maximizes α, situated around 78 K. In order to maintain a working temperature of 77-79 K

the column is inserted into a vacuum jacket and the condenser and the feed flow are cooled with liquid

nitrogen. The construction technique limits the number and type of sensors that can be installed on

the column, and so, the provided measured data.

A number of test runs have been done for this column, all running in total reflux. By choosing ex-

periment 2 [9] as a reference, we present the a priori information available for our model’s validation

in table 1, where HETP is the acronym for the Height Equivalent to the Theoretical Plate.

3

Modeling

Figure 2: Schematic representation of the 13C isotope separation unit

Table 1: MEASURED DATA IN EXPERIMENT RUN NO. 2

Running time (h) Concentration 13C (%) HETP (mm) Number of trays

96 2.71 26 164

3 Modeling

3.1 Introduction

The choice of a “general” model for a class of distillation columns is not easy and it is not feasible

to find one single model for all existent columns [11]. In the process systems community, different

approaches coexist and picking one of them is first of all subject to area of application. In the thesis,

we focus on first principles models for continuous operation mode to be used for control purposes.

3.2 Selecting a Suitable Model

Unfortunately, there are only heuristic guides for model selection based rather on rich experience

than systematic approach [12, 13]. Pearson [14] summarizes the most important “utility” measures:

approximation accuracy, physical interpretation, suitability for control and ease of development. Ap-

proximation accuracy refers to how well steady-state and dynamic model responses correspond to

the real process; physical interpretation describes the relationship of model and process variables;

suitability for control refers to adequate dynamics description and ease of development is understood

literally. Some first principles models meet these requirements very well. They are more accurate

4

3.3 Model Equations

Figure 3: Basic distillation column and general distillation stage

compared to linear models, easy to interpret and simpler ones are also easy to develop. The classi-

fication scheme in [13] outlines differences between the simpler and more complex first principles

models. Simpler models have been developed since computational resources became available, e.g.

[15], whereas more rigorous ones can be found in [16]. However, although more sophisticated nu-

merical solvers exist today, more rigorous models tend to cause numerical problems as they include

more nonlinearities and there could be not enough data to fit all parameters. Thus, we shall focus here

on the model which can be classified as an “intermediate” one and with which we could run stable

and fast simulations.

As an accuracy guideline for control, practitioners advice to match simulations’ steady-state to the

real plant operation within 5 %, i.e. flows, temperatures, pressures, compositions, levels and vessel

volumes must be similar to those of the process [17]. Once good estimates are obtained, the time

constants of the simulation and process will be similar and the validation of dynamic responses is

generally not necessary.

3.3 Model Equations

Figure 3 shows a common stage-distillation column for simulation studies with included parts:

condenser, rectifying section, feed tray, stripping section and reboiler. In order to model these parts

they are usually decomposed in single stages. Note that in packed columns in contrast to stage

columns, continuous contact between the phases prevails. However, both column types approximately

coincide in their numerical behavior and therefore can be depicted as stage columns [18].

The equations derived in [19] are based on several simplifying assumptions, the most important

5

3.3 Model Equations

are: holdup in vapor phase is neglected and for the liquid phase perfect mixing on a stage is assumed,

partial specific enthalpies of components are equal to the specific enthalpies of pure components

(ideality of the mixture). For the the i-th general equilibrium stage and the j-th component these

equations describe mass balance (2)–(4), liquid dynamics (5), vapor-liquid equilibrium (6) and energy

balance (7)

i = 1, . . . , N, j = 1, . . . , NC, (1)dMi

dt= Vi+1 − Vi + Li−1 − Li + Fi − Si, (2)

dMixi,jdt

= Vi+1yi+1,j − Viyi,j + Li−1xi−1,j

−Lixi,j + Fizi,j − Siwi,j, (3)nc∑j=1

xi,j = 1, (4)

Li = f(Mi, Vi+1), (5)

yi,j = Ki,jxi,j, (6)

VihVi = Vi+1hVi+1+ Li−1qLi−1

hLi−1− LiqLi

hLi

+FiqFihFi− Si+1qSi

hSi+Qi, (7)

where Mi is the total liquid holdup, Vi+1 and Vi are incoming and outcoming vapor flow rates, Li−1

and Li are incoming and outcoming liquid flow rates, Fi and Si are feed and side streams. x, y, z and

w denote molar or weight fractions. Furthermore, f is a function describing the liquid flow dynamics,

K is the so-called K-value, Qi is the heat input, q and h with the corresponding subscript represent

the thermal stream condition and heat of vaporization of the stream as a function of pressure and

concentration.

The feed tray is essentially a general stage without side streams.

The basic balance equations of the (total) condenser (8)–(12) correspond to those of a single

stage. The only difference is that it has an ingoing vapor stream V1 and two outgoing liquid streams,

the distillate D and the reflux L0, which are not independent. Furthermore, all the ingoing vapor is

condensed and thus it holds y1,j = x0,j

dM0

dt= V1 −D − L0, (8)

dM0x0,jdt

= V1y1,j −DyD,j − L0x0,j, (9)nc∑j=1

x0,j = 1, (10)

y0,j = K0,jx0,j, (11)

M0dqL0

dt= (1− qL0)V1 +

Q0

hL0

. (12)

The balance equations of the partial condenser(13)–(17) correspond to those of a single stage, as

presented earlier for the (total) condenser with the specification that y1,j 6= x0,j and ∆HV1 is the latent

6

3.4 Submodel-Library Implementation

heat of vaporization of the vapors that enter the condenser:

dM0

dt=

Q0

∆HV1

− L0, (13)

dM0x0,jdt

= V1y1,j −DyD,j − L0x0,j, (14)nc∑j=1

x0,j = 1, (15)

y0,j = K0,jx0,j, (16)V0L0T

dp

dt= V1 −

Q0

∆HV1

−D. (17)

The balance equations for the reboiler (N + 1-th stage) without external feed are

dMN+1

dt= LN −B − VN+1, (18)

dMN+1xN+1,j

dt= LNxN,j −BxN+1,j − VN+1yN+1,j, (19)

nc∑j=1

xN+1,j = 1, (20)

yN+1,j = KN+1,jxN+1,j, (21)

VN+1hVN+1= QN+1, (22)

where B is the bottom product stream and QN+1 is the heat flow rate from the reboiler.

Note that (5) can be a geometrically motivated relation, e.g. Francis weir formula [13], or its

linearization. The vapor-liquid equilibrium equation (6) is often chosen as yi,j = αjxi,j/∑nc

k=1 αkxi,k

with relative volatilities αj [20], but it also can be substituted by a distillation efficiency formula, e.g.

Murphree tray efficiency [21] or an advanced formula [22]. The main advantage of the simplified

energy balance equations, (7), (12) and (22), is their simplicity. The rigorous energy balance would

increase the model complexity and decrease its numerical robustness, whereas by removing the energy

balance equations, we would end up with a constant molar overflow model which might imply wrong

results in a control study [23]. Including a pressure drop instead of assuming a constant pressure

might be another important step to improve the model accuracy for columns with a large pressure

drop [24].

3.4 Submodel-Library Implementation

For our implementation we used the equation-based process modeling system gPROMSTM

. In

this environment, a library for modeling distillation processes was created. The library is separated in

categories, to ease the process of sub-model selection, into:

• MultiDistCol-Library-Reduced-313 with models for the component parts of the distillation

columns,

7

3.5 Model validation

• FlowControl-Library-313 with models especially for flow manipulation such as: pumps, valves,

splitters, tanks etc.,

• Control-Library-313 with models for the control elements.

The thermodynamical properties, e.g. enthalpies or boiling temperatures, are calculated by an

outsourced package and accessed by all models via Foreign Object interface. This allows us to de-

scribe properties of ideal and non-ideal mixtures and easily rewrite the models in reasonable cases.

Moreover, each model has ports to graphically construct composite models that speeds up model

aggregation.

3.5 Model validation

For validating the models in the library, two processes have been chosen: the ACBT process and

the 13C isotope separation process.

Both processes were modeled and the stationary profiles compared with the GAMS profiles for

the ACBT case, respectively with process measurements for the isotope separation process. In both

cases the fit was very good [25], for the ACBT case the root mean square error of the concentrations

for each separation stage was below 10−3. The fit for both processes proves firstly the quality of

the models and secondly their general character. However, the quality and generality of the models

is not concluded yet, more case studies with process measurements being required to validate these

statements.

4 Full-order nonlinear observer

4.1 Introduction

In order to apply effectively advanced control algorithms on distillation processes, the concen-

tration profiles are required. Because in practice, concentration measurements are available only in

exceptional cases, estimating these profiles with a state estimator from a low number of temperature

measurements represents an attractive solution.

As state estimators, Kalman filters are the most popular in industrial applications. However, with

the evolution of computational power it appeared the possibility to implement new state estimators,

numerically more complex, which previously was not achievable. Here, the observer presented by

Lang and Gilles [4] will be introduced and implemented on the ACBT process. It’s main advantages

lie in the few design parameters and the small number of temperature measurements required to ensure

observability.

8

4.2 State estimation

4.2 State estimation

Lang and Gilles [4] extend the full order non-linear observer presented by Zeitz [26] in which the

temperature difference between the process and model is corrected by the measurement in such a way

that the mass transfer of both regions coincide. This is done by influencing the mass transfer directly

and the general structure of the observer is represented in equation 23.

˙̂x = f(x̂, u) + α ◦ J ◦ (T − T̂ ), (23)

where: x̂ are the estimated liquid compositions, u the inputs, α the observer tuning parameter,

J the material flow , T the measured temperature, T̂ the estimated temperature and ◦ represents the

element wise multiplication.

The maximum number of elements (temperature measurements) required in order to guarantee

the columns’ observability is stated as number of components (NC) - 1.

The implementation of this observer has been cited in just a few studies and in none of these

studies was it used in closed loop, one of the reasons could come from the computational complexity

that a full-order nonlinear composition estimator introduces in a control system, problem that today

does not exist any more. Other full-order observers were cited by Shin et.al. [27] who includes in a

control structure, developed by one of the authors, a full-order observer modeled using wave models

and the performance is similar to the case when Kalman filters were used; the insensitivity to mea-

surement noise being stated as an advantage of the observer. Olanrewaju and Al-Arfaj [28] compare

a Luenberger observer with Kalman filters for a reactive distillation process with 2 reactants and 2

products. Between the advantages of the Luenberger observer design and implementation simplicity

are mentioned. However, the Kalman filter was found to be more robust to noisy measurements, erro-

neous initial estimates and model uncertainties. A moving horizon estimator extended to incorporate

a mode change detector and an operating switch estimator is used to estimate compositions for batch

and continuous distillation processes by Olanrewaju et.al. [29]. The algorithm proves very effective.

Considering these works and willing to see whether Lang and Gilles’s observer is suited for complex

distillation processes, a numerical model has been developed.

Lang and Gilles [4] suggest using at least one temperature measurement within the region of high

mass transfer in order to guarantee observability of the temperature profile. In order to find the region

of high mass transfer within the column the dynamic behavior after a change of the heat input has

to be examined. For this, simulations have been ran, on the ACBT process, with steps on the energy

input in the reboilers of the three columns. It were then selected the points with the greatest change

in the temperature and concentration profiles, corresponding to the regions with high mass transfer.

The observers’ tuning parameters were determined through experiments by simulation. The de-

termined best values correspond to: α=0.01 for the first column, α=0.006 for the second column and

α=0.001 for the third column.

To test the convergence speed of the observer a simulation test was ran under the following condi-

9

Control of distillation processes

tions. Starting from an operating point, different from the nominal one, the process and the observer

(with α=0) are left to run for 5000 s. At 5000 s the process and the observer are being brought

back to the nominal case and the α parameters set to the previously determined values. From these

simulations it can be seen that the observer responds faster than the process and converges to the

desired value for each temperature measurement. Therefore, we can state that this observer offers a

convenient solution for state estimation when in possession of a first principle model of the process.

5 Control of distillation processes

5.1 Introduction

The control of the distillation processes was approached from the plantwide perspective based on

Skogestad’s [5] concept. The plantwide perspective on control is needed not only, as many authors

point out, because of the changes in the way plants are being designed - with more heat integration,

recycles and less inventory, but also for plants without any heat integration [30] as chemical plants

consist of a string of units connected in series with one unit acting as disturbance to the next. Basically,

plantwide control follows a "top-down analysis to select controlled variables followed by a bottom-up

assignment and possibly design of the control loops". The complexity of the ACBT process with three

columns and recycle places this process as a perfect candidate to apply the plantwide strategy on. The

implementation steps will be presented in the next chapters.

5.2 Decomposition of the control on two layers

The scale and difficulty of designing control systems for complete plants suggest the decompo-

sition of the control problem into more manageable parts. The decomposition based on the control

objectives, is selected in the thesis, with respect to two objectives:

1. Stabilization of the process under the influence of perturbations.

2. Plant supervisation.

The two objectives will be managed by the stabilization control, respectively by the supervisory

control (figure 4). Here, the stabilization control represents the fast loop of the process; controllers

working on a shorter time scale. The supervisory control layer restricts the behavior of a plant such

that as much as possible of the given specifications are fulfilled. The controllers from this level are

generally model based controllers that set the references of the controllers from the lower level by

minimizing an objective function.

10

5.3 Control of the ACBT process

Figure 4: Decomposition of the control problem on layers

On the supervisory layer MPC is used to drive the process back to the reference trajectory with

respect to constraints, by minimizing an objective function as the one represented below:

J c =N∑k=1

((yk − yrefk

)TQ(yk − yrefk

)+(uk − urefk

)TR(uk − urefk

)+

+(δuk − δurefk

)TRδ

(δuk − δurefk

)) (24)

where yk are the controlled variables (outputs), uk the manipulated variables (inputs) and δuk =

uk−uk−1. The matrices Q, R and S are the corresponding weights on the outputs, inputs, respectively

on the δ inputs.

For the implementation of the control strategy we have chosen the DRTO Toolbox [6], an environ-

ment developed in the Chair of Process Systems Engineering from RWTH Aachen. The toolbox is a

real-time control framework that offers the possibility for the user to adapt its structure, by introducing

or removing modules to fit the control strategy that is to be implemented.

5.3 Control of the ACBT process

5.3.1 Introduction

Controlling the ACBT process will be done by separating the control problem in two layers: the

stabilizing layer and the supervising layer. To choose the stabilizing structure, different strategies

have been tested for each column and only the best ones were implemented.

For the supervising layer the LTVMPC, algorithm implemented in DRTO Toolbox, has been cho-

sen. At the end of each MPC time sample the controller will send either the reference value for the

controllers from the stabilizing layer or directly a control signal to the manipulated process variables.

The latter solution is used seldom and when the controlled loop is a slow one under the influence of

small disturbances.

11

5.3 Control of the ACBT process

5.3.2 Stabilizing layer

For the first column a one point control scheme has been chosen. Two point control could have

also been implemented, but because of stability reasons, under the influence of disturbances, one point

control has been preferred. The control structure is presented below in figure 5.

For the second column, the one with a partial condenser, 3 structures [31] were analyzed and

the one that provides smaller and more gradual changes in the vapor distillate flow rate was chosen

(figure 6). Small changes are desirable from a plantwide perspective because they introduce the

smallest disturbances to the downstream columns.

Figure 5: Control structure of the first column Figure 6: Control structure of the second column

For the third column, the classical 2 point (L/V) control has been implemented as in figure 7.

5.3.3 Supervisory layer

The starting point for plantwide control is to establish the number of degrees of freedom for

operation. Here we distinguish the dynamic degrees of freedom (for control) and the steady-state

degrees of freedom. The latter ones are used to be manipulated by the MPC algorithm and can

be determined by subtracting the number of variables with no steady state effects. After a degrees of

freedom analysis, 4 variables were selected to be implemented in the MPC strategy. These correspond

to the:

• flow of heating agent from the reboiler (Column 1)

• reference of the temperature regulator from the bottom of the column (Column 2)

• reference of the temperature regulator from the top of the column (Column 3)

12

Conclusions

• reference of the temperature regulator from the bottom of the column (Column 3)

Figure 7: Control structure of the third column

5.3.4 Results and discussions

For the ACBT process simulations were ran with a 4-by-4 control structure with constraints on

the inputs and on the controlled variables under the influence of a 5 % disturbance on the feed flow.

The controlled variables are the concentrations of the 4 components: acetone, chloroform, benzene

and toluene. Compared to the response of the process without feedback control and with the one with

only the regulatory layer implemented, the response of the system is much faster and the composition

specifications were respected, as espected.

6 Conclusions

In industry the control of distillation processes is done in 90 % of the applications with basic P, PI

or PID controllers and only the rest of 10 % is done with advanced control algorithms [1]. The slow

pace and the reserve of process engineers in implementing advanced control algorithms comes also

from a lack of general modeling and control frameworks, that would offer a user friendly environment

to test and validate control strategies before the actual implementation.

The doctoral thesis presented a modeling and control framework, that offers a unitary approach

of modeling and control of distillation processes. To validate the platform two processes have been

selected: the ACBT process and the 13C isotopic separation column.

13

Conclusions

In order to reduce the modeling time, a library with dynamic models was created that includes

models for the distillation column parts, for flow manipulation elements and for control elements.

Despite the good fit obtained for the two processes, a validation on a real industrial column with more

measurements is still necessary.

Because of the drawbacks that a centralized controller has, the decomposition of the control prob-

lem has been chosen. The plantwide approach offers the best option in this direction, where the

control is separated on two layers: a fast stabilizing layer and a slower one, also called supervisory

layer which sets the references for the lower level controllers. The results on the ACBT process

proved that the two layer control strategy, even though it requires more time and process insight in

the design phase, delivers better control in the presence of constraints and disturbances than a one,

regulatory, layer control strategy, both in terms of speed and robustness.

The results obtained for the ACBT process (strongly nonlinear 3 column process with recycle),

starting from modeling, state estimation and ending with the application of plantwide control strate-

gies, encourages further developments and brings the framework closer to industrial control applica-

tions.

14

References

References

[1] S. Agachi, Z. Nagy, V. Cristea, and A. Imre, Model Based Control. Case Studies in Process

Engineering. John Wiley and Sons, 2006.

[2] R. Taylor, “(Di)still modeling after all these years: a view of the state of the art,” IChemE Symp.

Series, vol. 152, pp. 1–20, 2006.

[3] R. Findeisen and F. Allgower, “An introduction to nonlinear model predictive control,” in 1st

Benelux Meeting on Systems and Control, 2002.

[4] L. Lang and E. Gilles, “Nonlinear observers for distillation columns,” Computers Chemical

Engineering, vol. 14, pp. 1297–1301, 1990.

[5] S. Skogestad, “Plantwide control: the search for the self-optimizing control structure,” Journal

of Process Control, vol. 10, pp. 487–507, 2000.

[6] D. van Hessem, “Implementation of the uppercaseINCOOP control and optimization architec-

ture – general documentation and tutorial on incoop software,” tech. rep., TU Delft, 2003.

[7] K. Kraemer, S. Kossack, and W. Marquardt, “Efficient optimization-based design of distillation

processes for homogenous azeotropic mixtures,” Industrial & Engineering Chemistry Research,

vol. 48(14), pp. 6749–6764, 2009.

[8] D. Y.-C. Thong and M. Jobson, “Multicomponent homogeneous azeotropic distillation 3. col-

umn sequence synthesis,” 2001.

[9] A. Radoi, M. Gligan, and A. Baldea, “Experimental plant for the 13C isotope enrichment based

on carbon-monoxide distillation at low temperatures,” Revista de Chimie, vol. 50 (3), 1999.

[10] D. Axente, M. Abrudean, and A. Baldea, Isotope Separation of 15N, 18O, 10B, 13C by Isotopic

Exchange. Casa Cartii de Stiinta, 1994.

[11] H. Z. Kister, Distillation - Operation. McGraw-Hill, 1989.

[12] E. F. Vogel, Practical Distillation Control, ch. Plantwide Process Control Simulation. Van

Nostrand Reinhold, 1992.

[13] V. G. Grassi, Practical Distillation Control, ch. Rigorous Modeling and Conventional Simula-

tion. Van Nostrand Reinhold, 1992.

[14] R. K. Pearson, “Selecting nonlinear model structures for computer control,” Journal of Process

Control, vol. 13, pp. 1–26, 2003.

[15] R. Levy, A. Foss, and E. Green, “Response modes of a binary distillation column,” Industrial &

Engineering Chemistry Fundamentals, vol. 8, pp. 765–776, 1969.

[16] R. Gani, C. A. Ruiz, and I. T. Cameron, “A generalized model for distillation columns-—i.

model description and applications,” Computers & Chemical Engineering, vol. 10 (3), pp. 181–

198, 1986.

[17] W. Luyben, ed., Practical Distillation Control. Van Nostrand Reinhold, 1992.

[18] W. Marquardt, Nichtlineare Wellenausbreitung – Ein Weg zu reduzierten dynamischen Modellen

von Stofftrennprozessen. PhD thesis, Universität Stuttgart, 1988.

15

References

[19] K. E. Häggblom, “Modeling of flow dynamics for control of distillation columns,” 1991.

[20] H. Z. Kister, Distillation Design. McGraw-Hill, 1992.

[21] E. V. Murphree, “Rectifying column calculations with particular reference to n component mix-

tures,” Industrial and Engineering Chemistry, vol. 17, no. 7, pp. 747–750, 1925.

[22] R. Taylor and R. Krishna, Multicomponent Mass Transfer. John Wiley & Sons, 1993.

[23] B. Wittgens and S. Skogestad, “Evaluation of dynamic models of distillation columns with em-

phasis on the initial response,” Modeling, Identification and Control, vol. 21, no. 2, pp. 83–103,

2000.

[24] J. G. Stichlmair and J. R. Fair, Distillation Principles and Practice. Wiley-VCH, 1998.

[25] I. Muntean, M. Stuckert, and M. Abrudean, “A general distillation modeling framework applied

to an isotopic distillation column,” Proc. of the 19th Mediterranean Conference on Control and

Automation, pp. 1150–1154, 2011.

[26] M. Zeitz, “Nichtlineare beobachter für chemische reaktoren,” Fortschrittsberichte VDI-Verlag,

vol. 8/27, 1977.

[27] J. Shin, H. Seo, M. Han, and S. Park, “A nonlinear profile observer using tray temperatures

for high-purity binary distillation column control,” Chemical & Engineering Science, vol. 55,

pp. 807–816, 2000.

[28] M. J. Olanrewaju and M. Al-Arfai, “Design and implementation of state observers in reactive

distillation,” Chemical Engineering Communications, vol. 195, pp. 267–292, 2008.

[29] M. J. Olanrewaju, B. Huang, and A. Afacan, “Online composition estimation and experiment

validation of distillation processes with switching dynamics,” Chemical Engineering Science,

vol. 65, pp. 1597–1608, 2010.

[30] T. Larsson and S. Skogestad, “Plantwide control—a review and a new design procedure,” Mod-

elling, Identification and Control, vol. 21(4), p. 209–240, 2000.

[31] W. L. Luyben, “Alternative control structures for distillation columns with partial condensers,”

Industrial & Engineering Chemistry Research, vol. 43, pp. 6416–6429, 2004.

16