9 Optimization of Process Operation

In this chapter, process optimization will be discussed. Afte r process control has been
improved and is robust and stable. the setpo ints ofthe control strategy can be opti-
mized. First process op timization is discussed in general and some ofthe p itfalls are
discussed. An illustrative example is g iven oflinear pro gramm ing and suggestions
are given how to perform non-lin ear optimization.

9.1 Introduction to Real-time Optimization

Once a plant has been built, there are usually many degrees of freedom that can be se-
lected in order to minimize the cost of operating the plant and maximizing the operat-
ing profit. To solve this problem successfully, it is neces sary to understand the theory
as well as the practice of optim ization. This chapter will help the reader to gain some
underst andin g of both aspects of an optim ization.

9.1.1 Optimization and its Benefits

In the field of mathematics and engineering, optimization is often referred to as the

selection of process variables in such a way that a pre-defined performan ce criterion
is minimized. In optimal control for example , it is quite common to minimize a quad-
ratic performance criterium, e.g. £2, where s is the deviation between a process vari-
able and its set point. This does not necessarily mean that the cost of operating the
process is minimized . To optimize a chemical or petrochemic al plant, we will have to
select the process parameters in such a way that the value of the plant profit function
is maximized. A different profit function may lead to another set of optimal process
parameter s.
In order to calculate the value of the profit function, a process model must be
available. The model must be capable of accurate predictions for a wide range of op-
eration s, e.g. normal operation, the use of different feedstocks, varying equipment
performance, etc. The model must also be robust so that it always gives an acceptable
solut ion. In order to define a meaningful performance criterium, the costs of all plant
inputs and outputs have to be known, such as the value of feedstocks, fuel, catalyst,
electricity, products and byproducts, etc.
Many optim ization problems in chemical engineerin g have an infinite number of
solutions. Optimization is concerned with selecting the best among this entire set. The
202 9 Optimization of Process Operation

engineer needs considerable insight as to what an appropriate performance criterion

would be and use his engineering judgement to determine whether a given set of op-
timal process parameters could be implemented. Expert system s could be useful in in-
terp reting some of the results of the optimi zation.
In plant operation, benefits arise from improved plant performance such as increased
yield ofvaluable products, reduced ener gy consumption, higher processing rates and
larger times between shutdowns.
Optimization is especially useful when constraints have to be monitored. The opti-
mum of a process is almost alwa ys found near one or more constraints. Selection of
process parameters in such a way that a safe operating distance from con straints is
maintained, while maximizing a profit function, can save millions ofEuros in operat-
ing costs.

9.1.2 Hierarchy of Optimization

Optimization can take place at different levels ranging from multi-plant optimization
to unit optimization. There is always one pre-requisite for succe ssful optimization:
The lower level control strategies must all perform well. This is quite obvious; if an
optimal set of process parameters is calculated, these set points must be implemented
successfully. Fig. 9.1 shows two possible approaches to optimization: centralized ver-
sus distributed optimization.

GENERALIZED OPTIMIZATION DISTRIBUTED OPTIMIZATION

UN IT OPTIMIZATION
GLOBAL UNIT OPTIMI ZATION MOD EL (COORDINATION MODEL)

ii I I
T
I II ! I I,
I
I1
!
•i ! " " •
; r-L---'---,
• :
• • •
AD VANC ED REGlJL ,.\TORY/CONSTRAI NT CONTROLS
LOCAL
OPTIMIZER
l.OCAL
OPTI M IZER I O~~~~~ER
BASIC CONTROLS DISTILLATION
! COMPRESSOR
!L-.,-----.--'
COLUMN REACTOR

ii I
PROCESS

.L I ; i .L ; :
i
ADVANCED REGULATORYICONSTRAINT CONTROl.S

BASIC CONTROI.S

PROCESS

Fig. 9.1. Centralized versus distributed optimization.

The advantage of the distributed or modular approach is that the total problem is bro-
ken down in smaller pieces and the problem can therefore be easier understood. An-
other advantage is that if the local optimizer for one unit is modified, the other opti-
mizers can still perform their tasks.
 9.1 Introduction to Real-time Optimization 203

A drawback ofthe distributed approach is the coordination ofall the local optimizers
by a global optimizer. This is not a trivial problem and needs extra care. If a number
oflocal optimizers each optim ize a unit, the result is not necessarily that the overall
profit function is maximized. Decomposition ofan optimizer can only be successfully
done when intermediate streams show little interaction in incremental prices and con-
straints .
Real time optimization will link the levels of scheduling and control in a particular
plant as shown in Fig. 9.2. There are a number ofsituations in which optimization can
become very attractive.

SCHEDULING

REAL-TIME
OPTIMIZATION

CONTROL

PROCESS
Fig. 9.2. Optimization as a link betweencontroland scheduling.

sales limited by production. Ifadditional products can be sold beyond current ca-
pacity, the economic justification for optimization is easy . Increased production
can often be realized by minor modification in operating conditions or changes in
operating costs.
sales limited by market. The opportunity for optimization in this case is to make
products more efficiently at a lower price .
large unit throughput. High production volumes offer an excellent opportunity
for increased profit since small changes in operating costs are multiplied by a
large number. An example of this situation are most ethylene plants .
high raw material and/or utility costs. Significant savings may be achieved by
reducing raw material consumption and/or energy consumption.
losses of valuable components in waste streams. A small reduction of the
concentration ofvaluable components in waste or recycle streams can contribute
significantly to the plant's operating profit.

As already mentioned in a previous section, there are a number of issues that need to
be addressed before tackling any optimization problem. A short review will be given
here.
204 9 Optimization of ProcessOperation

9.1.3 Issues to be Addressed in Optimization

Process model. The model must represent changes in the manipulated variables ofthe
process. These independent variables are often set points ofcontrollers, such as flow,
level and temperature controllers. The model will calculate the effect of changes in
the independent variables on the dependent variables. In most cases the models are
non-linear, steady state models.

Objective function. The purpose of the optimization is to minimize or maximize an

objective function or loss/profit function by manipulating the independent variables.
The objective function represents an economic model of the process. Only variable
effects should be taken into account and fix ed costs such as labour, overhead costs,
etc should be ignored. A typical objective function could have the following struc-
ture :
operating profit = product value - utility cos ts - [catalyst
replacement] -[longer term factor s, such as equipment (9.1)
service, overtime, etc]

Optimization algorithm. The optimization algorithm uses the unit model and the ob-
jective function to solve the problem. Many algorithms are available. Computational
requirements, effectiveness and robustness are the most important considerations.

Input data validation. Obtaining valid data from the process for the optimization
problem is a key concern. Some questions are: How fast to sample , how fast to aver-
age, what checks need to be done to test whether a signal is reasonable. If some data
is not available, how often should it be entered ? Erroneous or missing data must be
identified and other data must be reconciled to reduce the error and provide accurate
information about the current operating conditions.

Dynamic implementation. After the optimizer has calculated optimum values for the
degrees of freedom, their values should be implemented in the process. Checks must
be made to ensure that the result does not violate the current feasible operating win-
dow of the process. The actual operating window will always be different from the
operating window calculated from the model because of modelling errors .

Model updating. Process models are an approximation ofreality. As operating condi-

tions shift, some model parameters may have to be adjusted to describe the process
behaviour in the new operating point more accurately. New constraints may have to
be added to the model and others may become inactive. Model maintenance and man-
ual/automatic model updating are important issues . The main components of the op-
timization structure are summarized in Fig. 9.3.

Execution frequency. Because unit optimization uses steady state models, the execu-
tion frequency depends on the settling time of the process . The optimization interval
 9.1 Introduction to Real-time Optimization 205

must be longer than the settling time to ensure that the process is at a new steady
state. Fig. 9.4 shows the concept.
.· · · · · · · · · · · · · · · · · · · ··· · ··· · · · · ··· · · · · · · · · ·· · · · · · ·1··· · .
.

bad
data

bad
data

bad
data

Output
calculation

Fig. 9.3. Real-time optimization system.

optimization
interval
process
variable
y

- - - time
Fig. 9.4. Optimization with suitable execution interval.

Assume that changes are made to an independent variable u. The response of the
process variable y is such that in about two-th irds of the optimization interval a new
steady state is reached . If, however, the situation is as shown in Fig. 9.5, then the
process variable y never reaches steady state. A new set point is down loaded from
the optimizer before the proce ss variable has settled. This is an undesirable situation
and can in the worst case lead to process instability.
206 9 Optimization of ProcessOperation

process
variable
y

Fig. 9.5. Optimization using inappropriately selectedoptimization interval.

9.1.4 Degrees of Freedom Selection for Optimization

In this section degrees of freedom selection will be illustrated using three different
examples: boiler optimization, distillation tower optimization and ethylene plant op-
timization.
The degrees of freedom are generally those process variables that have not been
used in a basic or advanced control strategy and are therefore available to optimise
process operation.
In a boiler it is desirable to optimize the fuel/air ratio such that the boiler efficiency
is maximized. On the other hand environmental regulations encourage operation un-
der fuel-rich conditions and lower temperatures to reduce emissions of nitrogen ox-
ides. Unfortunately these operating conditions decrease the efficiency due to un-
burned fuel that in tum results in increased hydrocarbon emissions.
Fig. 9.6 illustrates the trade-off between efficiency and emissions, efficiency can
be maximized and emissions should be minimized, resulting in a trade-off between
two objectives.
Another simple example of optimization is found in distillation tower operation.
Operators tend to over reflux a tower since it will be easier to produce "on-spec" ma-
terial. But over refluxing wastes energy since at larger reflux rate more liquid has to
be boiled up. Within the tower constraints (flooding, weeping, etc) there will be an
optimal reflux at which steam usage is minimal in the light of a target overhead com-
position.
Table 9.1 shows the degrees of freedom in a typical ethylene plant, grouped ac-
cording to operating unit. Because of its complexity, numerous degrees of freedom
exist and there is ample room for optimisation.
 9.1 Introducti on to Real-t ime Optim ization 207

Efficiency Emissions

- NO

o 2
air/fuel ratio
Fig. 9.6. Efficiency and emissions for a typical boiler

Table 9.1 Variables for on line optimization in an ethylene plant.

Cracking heaters, for each heater: Depropanizer

Feed rate Propane in bottoms
Steam/ HC ratio Reflux rate
Severity Pressure
Charge gas compressor Propylene fractionator
Suction pressure Propylene impurity level
Reflux rate
Pressure
Demethanizer Debutanizer
Methane in bottoms Reflux rate
Reflux rate Pressure
Pressure Refrigeration
Deethanizer Ethylene refrigeration
Ethane in bottoms Compressor suction pressure
Reflux rate Propylene refrigeration
Pressure Compressor suction pressure
Ethylene fractionator Acetylene/ MAP converter
Ethylene impurity level Temperatures
Reflux rate
Pressure

9.1.5 Procedure for Solving Optimization Problems

For the solution of any optim ization problem, the follo wing procedure should be fol-
lowed:
analyse the process and mak e a list of all the process variables
208 9 Optimizationof Process Operation

define the economic objective function in terms ofthe above variables. Pay spe-
cial attention to the value of intermediate flows of material (this is a difficult is-
sue!)
develop a process model and define all process and equipment constraints. Use
mass and energy balances and empirical equations. Identify the dependent vari-
ables and determine the number of degrees of freedom.
if the problem formulation is too large , break it up into easier to handle parts or
simplify the model.
select a suitable optimization technique.
check the answers and check the sensitivity ofthe result with respect to changes
in the coefficients in the objective function, changes in independent variables and
the assumptions.

For a successful optimization it is essential to keep the model and objective function
as simple as possible without loosing the essence of the problem.

9.1.6 Problems in Optimization

Some ofthe early attempts to process optimization were not always successful. Some
of the reasons that were valid in those days are still valid today.

Maintenance requirements. It is alway s easy to justify adding new variables to real

time optimization problems. In a slightly different situation, other variables may be-
come important. There are also many constraints on process operation. Including
every possible variable and every possible constraint will result in a model that will
be too cumbersome to maintain. In addition, complicated models require sophisticated
maintenance and well -trained personnel. The more complicated the model the higher
the required skill level will be. Several early projects were successful initially, but
failed eventually because of inadequate number of support people. Because of the
high cost for maintenance of the models, they should be as simple as possible and
only represent the essential effects and relevant constraints. The practical relevant
question is always: how detailed should the model be?

Input data reliability. Process measurements are sometimes unavailable for large pe-
riods oftime. If during such a period the optimizer runs, it may have invalid inputs for
one or more process variables. The solution it will calculate may therefore be infeasi-
ble. Early optimization projects suffered much from unreliable instrumentation and
poor or unavailable measurements. Failure to provide sufficient validity checks on the
input data may lead to significant problems.
The lack of key measurements may necessitate interpolation ofunavailable meas-
urements between updates using a good process and disturbance model with an up-
date every time a measurement shou ld become available.

Changing unit conditions. It often happens that certain parts of the plant are not
available for optimization because ofequipment problems or maintenance. This sup-
 9.2 Model Building 209

ports the need for modular optimization even thou gh it might be more difficult in
some cases becau se of the integration of all optimization modules. The optimization
for that section of the plant that is currentl y not on stream can then be switched off.

Computational requirements. Some ofthe early systems were not powerful enough
to solve the opt imization problem within a reasonable time. Even though systems are
considerabl y more powerful, attent ion still has to be paid to make models as efficient
as possible in order to achieve the shortest possible computation time.

Human interface. If the operator does not understand what the purpose of an opti -
mizer is, he will have an excu se to tum it offand blame it for making erratic moves to
process variables. The human interface for some of the early projects was not always
considered important and did not make the results easy to understand. This problem is
still relevant today and special attention should be paid to designing an excellent op-
erator interface.

Lack offeedback on profit. Ifthere is no feedback on the result ofthe optimization,

it will be difficult to improve the objective function. In fact, there will be little incen -
tive for continued opt imization .

9.2 Mode l Building

As already discussed in chapt er one, modell ing is valuable because it is a mathemat i-

cal formulation ofreality and helps avoid repetitive experimentation and observations.
It also allows us to perform calcul ations how to operate the plant in the best possible
way. However, the time and cost savings by using the model should be weighed
against the cost of developin g the model. There are a number of issues that should be
addressed :

can the process be modelled by using first principles?

what is the required model accur acy and what is its impact on the final use ofthe
model?
are measurements available for model verification?
can the process model be decomposed into smaller subsystems which are easier
to analyse? This is especially important for trouble shooting.
should the mode l be linear or non-linear? Usually the model must be valid over a
wide range of operating conditions, often resulting in a non-linear model.
should the model be a distributed or lumped parameter model? The latter types of
models do not use a geometric coordinate and are there fore be simpler.
210 9 Optimization of Process Operation

9.2.1 Phases in Model Development

Model building can be divided into a number ofsteps: problem definition, preliminary
and detailed design analysis, evaluation and interpretation. Model building is an itera-
tive procedure. Figure 9.7 summarizes the activities to be carried out. For a more
comprehensive treatment of model building, the user should consult the literature
[Edgar and Himmelblau, 1988]. A short review will be given :

Problem definition. In this phase the user has to define the problem that has to be
solved. The degree of model accuracy and use of the model should be determined. It
is also necessary to evaluate the model structure and complexity:

identify the number of independent variables

identify the equations required to describe the process
determine the number of unknown parameters

Comput er simulation.
sotiw are development

MODEL
DESIGN

MODEL
EVALUATION

Fig. 9.7 Important activities in model building.

Model building is an iterative process as shown in Fig. 9.7. Before starting model de-
velopment it is necessary to formulate an economic justification

Model design. Activities in the design phase include the following :

specification of the information content

 9.2 Model Building 211

formulation of the mathematical description of the model

description of the programming logic and algorithms necessary to develop and
use the model
simulation of the model

First the input and output variables should be defined. In addition the assumptions and
limitations ofthe model have to be evaluated. The computer hardware has to be speci-
fied and selected, program logic and flowsheets have to be developed and program
modules and their structural relationships have to be defined.

Model evaluation. This phase is the test of the model. The integrity of the entire
model according to evaluation criteria that have been defined should be tested. Also
the sensitivity ofmodel inputs and parameters should be tested to ensure that relation-
ships are meaningful. Use actual plant data to test the model. This step is referred to
as diagnostic checking and may entail statistical analysis of the fitted parameters
[Himmelblau, 1970; Box et ai, 1978].
Model validation consists of three parts: validation oflogic, validation of model as-
sumptions and validation of model behaviour. Comparison with historical input-
output data or comparison with pilot plant is required.

9.2.2 Fitting Functions to Empirical Data

A model relates the dependent variables ofthe process y to the independent variables
u. By using experimental data we can determine the form of the model and subse-
quently estimate the parameters.
As discussed in chapter one, models can be written in several mathematical forms
[van Lith, 2002] . Fig. 9.8 shows some possibilities. This section emphasizes estimat-
ing coefficients in simple algebraic type models.

Algebraic Differential Difference

equations equations equations
(continuous)

Partial Ordinary
differential differential
equations equations

Fig. 9.8 Mathematical form of a model.

Selection of the form of a model requires expert knowledge to recognize how the
changes in the data can be explained by a mathematical description. Optimization
212 9 Optimization of ProcessOperation

methods can help in the model structure selection as well as in the estimation of the
unknown coefficients. The best model shows the least error between the actual proc-
ess data and the prediction. Some empirical models could be:
Y= ao + al UI + a2 U2 + ...
Y= ao + all uf + al 2 UI U2 + ...
(9.2)
I
y/u = 2
a o +a I 5+a 25

The first equation is linear; the second equation is linear in the coefficients but
nonlinear in the variables and the third equation is nonlinear in the coefficients.
When the model is linear, the coefficients can be estimated using linear regression;
in other cases non-linear regression is required.
In fitting an empirical or theoretical model one should keep in mind that the number
of data points must be greater than the number of model parameters.
For example with four data points one can estimate a maximum offour coefficients
(Fig . 9.9). A straight line correlation for the four points might be adequate, but it
could be that the data can be fitted better using a quadratic model.

.: / .../ ....•
....
~ ,

-+ Quadratic

Fig. 9.9 Linear versus quadraticfit for four data points.

One should use physical insight and common sense in fitting data. In Fig. 9.9 the
data shows a minimum. Can this minimum be expected from physical insight or ex-
perience? If there is no information available to give insight into the structure, the
simplest model should be used .
If n is the number ofdata points and m the number ofparameters in the model then
one should collect enough data points such that n > m. A mathematical technique such
as e.g. least squares can then be used to minimize the sum ofthe square ofthe errors.
 9.2 Model Building 213

9.2.3 The Least Squares Method

The least squares method was already discussed in section 3.5 for the estimation of
model parameters. The discussion in this chapter will be focussed on solving sets of
linear equations. Although the problem is in essence the same , the problem formula-
tion is slightly different.
For each of n data points, an error t j can be defined as the difference between the ob-
servation j j . j =1,2, ... n, and the predicted model response Ymi u):

Y j - Ym .j =Cj j = l,....n (9.3)

The independent variables U are assumed to be known exactly. The overall sum ofthe
squared of the errors forms the basis for an objective function:
n

f= Lf:] (9.4)
j =1

A linear model Y = Po + fJl U will be used as an example to illustrate the least squares
method. The objective function becomes then:
n n

f = I:(Y j - Ym.Y = I:(Y j -P O-PIU j)2 (9.5)

j=1 j =1

There are two unknown coefficients ~o and ~I' and n known pairs ofdata points Yj and
Uj. We like to minimizefwith respect to ~o and ~ I. The first partial derivatives off
have to be equal to zero :

(9.6)

Rearrangement leads to a set of linear equations in two unknowns:

(9.7)

The summation of ~o is equal to n~o, thus the equations can be written as:
n n

nfJo + fJ,Luj = LYj

j=1 j =l
(9.8)
fJoIu j + fJlIu/ = IU jYj
j=1 j =1 j =1
214 9 Optimization of ProcessOperation

These linear equations can be solved quite easily for ~o and ~l'

Example

Fit the model y = ~o + ~I U to the following data (y is the measured response and U the
independent variable:

U Y
o 0
I 4
2 8
3 12
4 16

The computations will need the following information:

LUj = 10
LYj=40

which results in:

SP o + 10PI = 40
(9.9)
10Po + 30PI = 120
from which ~o = 0 and ~1 = 4, so that the model becomes y = 4u where y is the pre-
dicted value for a given input u.
The procedure will now be extended to a model which is linear in the coefficients:
m

y ='LfJ;u;, U o =1 (9.10)
;=0

Equation (9.10) is linear with respect to ~j but Uj can also be non-linear. However, the
values of Uj will be substituted before solving for ~j, hence the functional form of Ui'
does not matter. For example if the model is quadratic:
y = Po + P1u+ P2U2 (9.11)
the basis functions become:

(9.12)

thus maintaining the general structure ofEq. (9.10). Substitution ofEq. (9.10) into the
objective function (9.4) results in:
 9.2 Model Building 215

f= }; (Y j- Ym,)2
j=O
(9.13)

The independent variables are identified by a double subscript, i (0, ..., m) for the in-
dependent variables and j (I, ..., n) for the sequence of data points.
In order to determine the unknown variables differentiatef with respect to Po, ... Pm
and equate the derivative to zero. This results in (m+ I) equations for p:

13o'I UO/ + 131'I UO jU\ j + ... + 13m'I UO jUmj = 'IY jUOj

j=1 j =\ }= I j=\

130'IU ljUOj + 13\ 'IUlj 2 + ... + 13m I,U ljUmj = I,YjU 1j (9.14)
~ ~ ~ ~
n n n n
130 LUmjU Oj + 13\ LUmjU ij + ... + 13m LUmj 2 = LYjU mj
j=\ j=\ j=\ j=1
Note the symmetry ofthe summation terms in Uij. This set of(m+ I) equations can eas-
ily be solved on a computer. Equation (9.14) can be expressed in more compact form
using matrix notation.

flo y\
fll y=
Y2
fl = (9.15)

fln Yn

UII UI 2 Ulm

U21 U2 2 U2m
U= (9.16)

U nl Un 2 ... U nm

Equation (9.14) can then be written as :

U TU13=UT y (9.17)

from which the solution follows:

13 =(U T U r IU T y (9.18)
216 9 Optimization of ProcessOperation

9.3 The Objective Function

The objective function defines the function that should be maximized or minimized.
In terms ofplant operation it is usually the profit function which consists ofa number
of terms, representing income and expenses.
Equation (9.1) is an objective function in its most general and abstract form. In de-
fining a more specific objective function, one should realize some ofthe mathematical
aspects ofthe optimization calculation. Most mathematical techniques handle discon-
tinuities poorly, hence they should be avoided. Objective functions that contain ex-
pressions such as log x and l/x should be modified to avoid discontinuity. Often it is
possible to approximate these functions by polynomials in x.

9.3.1 Function Extrema

Most properly defined objective functions will exhibit extrema. There are, however,
different types of extrema as shown in Fig. 9.10.

f(xl f(xl flx)

smooth discontinuous nondiffcrcntiable
function function function

x x x

f(x) unimodal function flx) Multimodal function

one minimum multiple minima

x x
Fig. 9.10 Types of extrema.

It is evident that in optimization problems, smooth unimodal functions are preferred.

If there is a problem with finding a solution to an optimization problem, it may be
worthwhile to get a better insight in the shape of the objective function.

9.3.2 Conditions for an Extremum

The objective in optimization is to find the minimum or maximum of an objective

functionf(x) where x is a vector representing one or more variables.