Professional Documents
Culture Documents
527234
527234
527234
t\.)
c,
t:;'
....
-.
;::::
l::l
....
c
:::l
('l
c
E'"
-
....
...
:::l
l::l..
0'\
("')
C
:::l
....
::::::
<'l:>
"'I
to the model output. The constraint handler uses a process model to simulate process
responses over some horizon and to detect whether constraints at inputs, outputs or
states are violated. If so it will adjust signals.
5.2 Internal Model based Control for a Distillation Column
In figure 5.2 the distillation process and the controller are depicted. Of the primary
process control level only the PID controllers that receive setpoint values of the
MIMO controller are shown.
Inputs of the controller are:
top purity
bottom purity
feedflow
Outputs of the controller:
setpoint for the reboiler steamflow
setpoint for the reflux flow
The controller architecture is not listed here because it is confidential informa-
tion. The general structure of the controller is given in figure 5.1. The controller
contains only two models. Apart from the process (including the primary control
level) transfer model, the feedflow is the input of the feedforward disturbance model.
The feedforward setpoint compensator is not used and the feedback controller is a
Proportional action only. A limiter limits the control signals to the process and the
controller uses de-filtering. Defiltering is applicated when both the filtered and the
un-filtered version of a signal should influence the steer signal. The models are State
Space versions.
5.3 Implementation on TDC 3000
5.3.1 Requirements
The MIMO controller acts as a unit process controller. As depicted in figure 4.2 the
unit process control level is covered by the (A)PM. The IMC controller was compact
(State Space) enough to implement it in the PM (Process Module Data Point). The
implementaton of an explicit constraint handler was not part of the project. If it had
been or will become, it ought to be implemented on the AM (unit optimization level)
where there is enough space even to deal with a number of complex soft constraints.
The goal of the project was not to prove that the controller performed well
(control the process) but that it could be implemented easily on IDC 3000 (A)PM.
This is the reason why the models used in the controller could do as the process
simulator during the final tests. This raises one withdrawal: The MIMO controller is
based on models of the primary process control level + the distillation process. If the
models used in the MIMO controller function as simulators, the controller should be
attached directly to the simulator. In this case the operator can only enter the desired
top and bottom qualities and simulate the feedflow through a setpoint value. To
increase the operator's power, the same PIDs as in the primary process level and
37
w
co
TDC-3000
Implementation .
blockscheme
IPCOS IMC
Operator
settings
field inputs
digital)
>... ".3'-' eywell Amsterdam
..... 11-92
digital
CL(JLOCK Col_Mirna
:;0,
;-<euoilel
rlf) 0"'1
Umk(l)
1.:Sp(2) 1------1..
C,JS Aulc
Usp(1) 1------1..
7
",
(cJS Aulo
Ilr,I-,(l: 1--------------------------'
PIO ",'!
Ulilkl7J JI
Q V 302
FOe"
F"e:J Fl ',w
OIA30:-
ToC' Qualil y
011,30'
eu.]:;',
'Jrli.
L:_:__PI_[l__
---11-1 YIl)
I" DAS PV ., CALC (,1 (SPI Ysk(1)
L"C NTV, p. ", YI<(2)
" {iA:; PV (A' r (", (:,P) Ysi2)
C, 301
t"'"
c
.g
s:::i
'E'
.,
-
;::..
(J
w
s
-
;::..
t...l
c::.
c::.
Ilrll..d2:
'-----------------------------4-1
AI'I'IIOVAL
1 , ..
Irn,s
1;11-
...... :j,'.'.( "CI'
:)IA[;I1AM : ClOP
:J' _,"'1" " 10.\-1,
iNTERlljAL M:JDEL : Of\: T LER
:,
I '1.0.
,--
'-1111:) FOP DCTIL LAT10:-1 LOLUI'N
f',:, ... '
: -'l:' 79 Ion JI'I
.
All
I I I I I I I
r'""f 'Jr, 372201 .... ,
Honeyweil
lNOLJSTR1Al AliTOMr"::N
PM300 \' -,I,
Program structure
Sequence Col_Mimo(PM; Point Col_Mimo)
External definitions and references
Local Variable definitions
Local Constant definitions
Phase Init
Define variables with a constant value (singles and arrays)
Phase Control
Step Read Val
Read Quality PV's
Read Quality SP's
Read Feedflow PV
Read Reflux and Reboiler PV's
Calculate Fs filter
Step Proc mod
Multiply Am_Xmk = PJn * Xmk
Multiply Bm_Xmk = 8m * Xmk
Add XJm< = Am_Xmk + Bm_Xmk
Multiply Cm_Xmk = em * Xmk
Multiply Om Umk = Om * Umk
Add Ymk = Cm Xmk + Om Umk
- -
Step Dist_mod
Multiply Ad_Xdk = Ad * Xdk
Multiply Bd Udk = Bd * Udk
Add Xdk = Ad Xdk + Bd Udk
- -
Multiply Cd_Xdk = Cd * Xdk
Multiply od Udk = Dd * Udk
Add Ydk = Cd_Xdk + od_Udk
Calculate Fr filter
Calculate Reflux and Reboiler SP's
Goto Phase Control
End Mimo Col
Figure 5.5 Sequential Program Structure for MIMO Process Module Data Point
40
cI) linqLd :3 t er'
..._-- - ---II
-1
., _.. , , ' ,
h. l' 1'.J ,.,
, -__.._..--.--:-,,---..1
:30.110 I
'.. :.1 I
. -- I .111'''''''''''''''''''')
I .Refluxtank
I
l....lIlol4--fo!-r-r-rt ======== I;';' r-(O} Cp b ,i
I
L.U.__ __ _Jp. _ S..
I T - ,- I" "., .. j I I 0. f
l
I.U 1..1 ..'; 1.1 I ..J I. '.,
I I'
L
I
r
-lo........ -+_. .J .-..
--- ---- --.-.--..-.- ------( iJ ", 1 f F"1 ';1
b_l_
1
1_'.... 1_e_r'__...... jIalll- __.-l._ .
Condensate Bottom product
i
r-----
i .'-.
i
'F'
i, "0"
"ToO-
I PCO': :'I I 1,1 Ci
IMC DISfILLHTION CJLUMN
three delays were interconnected between the controller and simulator. In this
configuration, disturbances can be introduced by changing the tuning constants of the
PID controllers. The limiting of the setpoints requires some logic as well.
It was only practical to implement the simulator in the same PM. Eventually
the required system configuration was:
1 x User Station
1 x Local Control Network
1 x Network Interface Module
1 x Universal Control Network
1 x Process Manager, release 300
1 x alarm printer
The resulting architecture does not contain a unit optimization level to provide the
MIMO controller with setpoints for top and bottom quality. Instead they are entered
directly by the operator from the US using the setpoints of the interconnected PIDs.
This also holds for the simulated reflux- and reboiler valves. When the PID control-
lers are in automatic mode, the operator can enter setpoint values for the PID
controllers during the "start-up" of the process. After this, the controllers are switched
to cascade mode (normal operation). In figure 5.3 the IDC 3000 implementation
blockscheme for the IPCOS IMC is shown. In figure 5.4 the resulting loop diagram
for the PM 300 is depicted.
The Sequential Process Control structure of the controller in a PMDP is given in
figure 5.5. In "Step Proc_mod" the state and output vectors of the process model are
updated, in "Step Dist_mod" the state and output vectors of the disturbance model
are updated. This updating requires matrix multiplication for which no standard built-
in algorithm is available right now in the PM (in the AM it is). In this case all matrix-
multiplications are written out in full CL which makes this implementation very
specific. In figure 5.6 the operator's view on the controlled process is shown.
5.3.2 Results
In table 5.1 the available and required performance for the implementation of the
MIMO controller in the (A)PM 300 is given.
42
Table 5.1 Available and required performance of the (A)PM 300 for implementation
of the MIMO controller
# Memory Units # Processing units # Variables
System capacity (available)
APM300 10000 1600 2048+80*n
PM 300 3200 1600 2048+8O*n
Control Scheme (required)
-Basic Control 70 70 at 1 s scan
-MIMO Control 100 Max. 150 at 10 s scan 300
Aver. 15 at 10 s scan
-SP limits processing
(estimation) 35 Max. 20 at 10 s scan 25
Aver. 2 at 10 s scan
Total amount
required 205 for APM 2.1 % Max. 240 15% 325 < 16 %
205 for PM 6.5 % Aver. 90 5.6 %
The required amount of CPU is expressed in so called processing units. The scanning
frequency of the MIMO controller can be specified and is typically 10 times slower
than the scanning of the connected primary controllers. The memory size is denoted
by the number of Memory Units. The number of Variables depends on the number of
Process Module Data Points (n). Each configured PMDP has 80 local variables.
5.3.3 Conclusions
In table 5.1 the most important constraints are the number of variables and the
maximum number of processing units required. These values determine the number
of control loops that can be implemented in the (A)PM. As to this number there is
no difference between APM and PM. In this implementation 6 MIMO loops including
additional requirements could be implemented but in practice one per (A)PM would
do.
The number of variables can be decreased and robustness can be improved by
declaring the parameters that do not change (contents of State Space matrices, scaling
factors) to be constants. This would mean a variable reduction of at least 70 %.
Drawbacks of this operation would be that constants could not be changed and
visualized during runtime, the use of generic subroutines is impossible (manipulations
index impossible) and declaration in both controller and simulator is necessary. The
operation would make it possible to implement 8 MIMO controllers in one (A)PM.
Because only one process-unit MIMO controller is required in the APM, its complex-
ity can be increased and a constraint handler could be added.
If the PID controller is out of cascade mode, the MIMO controller sequence
"holds" while the process simulator continues. Of course, this causes a large error
between MIMO controller output and process simulator output and consequently,
when switching back in cascade mode, it takes a long time to get back the process
simulator PVs according the SP values.
Mter all we may conclude that the (A)PM allows MIMO controller implemen-
tation without any problems as far as (A)PM performance is concerned (takes only a
43
few percents of total). The MIMO controller perfectly copes with the primary
controllers in the (A)PM. In particular as far as bumpless transfer is concerned after
switching back to cascade (MIMO) control. The operator need not worry and can
switch to cascade without disturbing the process. The MIMO operator interface
behaves like a PID operator interface and needs the same approach. Once the MIMO
controller design is available it is only a matter of man-weeks to get it implemented
in (A)PM, tested and documented. If an existing plant has been equipped with a DCS
like IDC 3000 then there is actually no problem to move from primary control to
MIMO control. The only real task is to get the right MIMO controller designed.
Companies like Setpoint IPCOS can provide this kind of service. No particular
managerial problems (planning, financial, responsibility of human resources) were
encountered during the MIMO pilot project.
44
6 Software and Hardware
6.1 Developments in Software and Hardware
The level of technology is determined by the level of hardware and software separate-
ly and their integration. The technology is up-to-date if the software uses the capabil-
ities of the up-to-date hardware to the full extent.
There are four criteria for technology [7]:
1 performance
2 pnce
3 reliability
4 programmability and maintenance
The first three are hardware criteria, the fourth is a software criterion.
Nowadays the price of the hardware has become a minor issue. The price of a
process automation project is mainly determined by the hours that the highly skilled
project team spent.
Reliability is guaranteed by the use of high quality, proven elements. Some
elements are duplicated or even triplicated and as ultimate rescue, there is an
emergency handling system.
In section 6.1.1 performance will be dealt with more explicitly.
6.1.1 Performance
The performance of a hardware configuration (CPU, ALU, memory, bus) depends on
its accuracy, memory, speed and database access.
The accuracy is determined by the number of bits in the "currency" of the
Central Processing Unit and bus. This currency is called "word". Typical lengths: 8 bits
(= 1 byte), 16 bits or 24 bits.
The memory is characterized by its size expressed in the number of words that
can be contained.
The speed of the CPU depends on its clock, the number of bits in a word and
the way operations are implemented.
The frequency of the clock determines the duration of one cycle. All arithmetic
operations executed by the CPU take a number of cycles. Typical values; logic
operation: 1..2 cycles; addition/subtraction: 2 cycles; multiplication: 4..8 cycles (mean
value: 5); Division: 8..16 cycles (mean value: 10). The number of cycles that an
operation takes is determined by the way the processor is programmed.
The speed of processors has increased over time through the availability of
hardware allowing higher clock frequencies, smarter programming and increased
lengths of words. A cascade architecture of more than 1 servers, one (coprocessor) of
them receiving the overflow of others, further increases speed. In table 6.1 the INTEL
45
family is given with clock speeds.
Table 6.1 CPU clock frequencies [12]
CPO (INTEL)
8088
80286
8086
80286
80386
80486
80586
Clock Frequencies (MHz)
4.77
6/8
8
10
16/20
25 / 33 / 50 / 66
33 / 50 / 66 / 100
Whilst the individual processors are being optimized, the speed of the overall system
(CPU + ALU + memory + I/O) is determined by a bottleneck like the data-base
access. Physical constraints are envisaged that makes research looking for less
constrained alternatives.
6.1.2 Programming Languages
The software part of a specific application consists of algorithms that are imple-
mented in a specific programming language. Such an implementation is called a
program. Apart from the quality of the algorithm the language [12] is an important
factor in the quality of the program.
In hardware newly developed generations of computers replace old ones. In
computer languages, developments are being divided in generations as well but there
is a difference. As for the Hierarchical Control Structure, newly developed levels or
generations are added to existing ones, making a higher level of abstraction possible.
Programming languages can be divided into 5 generations:
The first generation is machine language. Instructions are series of bits that are
entered by the programmer.
The second generation is assembler language. The programmer enters mne-
monic abbreviations of the instructions. There is still a 1 to 1 relationship between
the program and the instructions for the computer. The assembler translates the
source code to the object code (machine language). A program in assembler language
is performed fast, efficiently using the processor. Assembler languages are dedicated
to a specific processor using its strengths. A higher language (>2rd generation) can
never reach this.
Procedural languages belong to the third generation. Each command means
several instructions for the computer. Procedural reflects on the fact that the program
must contain data and procedures to provide action in all states of the application. A
compiler translates the program to the object code which is saved and executed. If the
translation and execution are performed for each line separately, it is called an
interpreter. Procedural languages offer machine independency. Different languages
were developed for different applications. COBOL for business and administration,
46
FORTRAN and MATLAB for science, Turbo Pascal for education, BASIC for PC
environments, C for packages on PCs being powerful, flexible and fast.
Languages of the fourth generation want the user to specify what has to be
done rather than how in order to decrease the step between problem definition and
program. Examples for PC environments: Oracle, dBASE.
The fifth generation programming languages have powerful commands and data
structures that can contain knowledge in order to solve problems. Usp and Prolog are
languages in which knowledge can be structured in different ways. In Expert-System-
Shells the structures in which the knowledge must be captured together with the
methods to reason are fixed.
6.2 Hardware and Software requirements for the Hierarchical Control System
In this section hardware and software requirements will be treated for the levels of
the Hierarchical Control System of figure 3.2.
If the controller in the primary process control level is a SISO PID then output
processing (given the noninteractive scheme, equation A, section 4.3.1 and the mean
values for operation cycles in section 6.1.1) based on mean values costs 41 cycles.
Choose one cycle to take 1 micro s (equals 1 MHz clock frequency) then approxi-
mately 250 PVs can be sampled 100 times per second. Because a normal value for the
required sample frequency in chemical processes is 1 Hz, one should think that there
is a large overhead of speed which can be used by scanning a lot of PVs. However,
the collection of PVs (AD/DA) is not the only function of the processors of the
(A)PM: signal pre-processing, alarm functions, PV checking, ranging, etc. To guaran-
tee the necessary level of reliability, at least 50% of the CPU capacity must be used
for diagnostics. As was shown in figure 4.2, the unit process control functions
(sequential process control, see chapter 5) are performed by the same (A)PM CPUs
as well. Conclusion: the available CPU capacity is the bottle neck. The number of
variables or local memory at disposal suffices in (A)PM.
The CPU capacity can be increased through more economic use of the
applicated ones (software problem) or by replacing them by more recent and
advanced CPUs. In the latter case a lot of software has to be rewritten.
The identification and controller design part of the HMPC package is only partially
LeN resident (figure 4.8). The moment that an on-line identification (optimization of
model parameters of which the structure was determined off-line) and tuning of
controller parameters can be performed within the control system (all parts are LCN
available) and the whole package can be engineered quickly, demands a minimum of
expertise, can be applicable to a range of processes and is fully supported by its
supplier, then MIMO control will have become an accepted block-structured control-
ler within TOC 3000 like the PID controller nowadays. The technology to incorporate
the package is already available yet not in IDC 3000.
The algorithms within the HMPC package were based on technology from the
late seventies / early eighties that has been improved by now. Now it is possible to
produce more compact models, decouple the control and optimization problem, and
allow for robust control. Conclusion: the software can be improved as well.
47
The HMPC documentation described that the non-parametric description of
the process was saved in the AM. However, it did not treat where the controller
actions were calculated. Here it is assumed that the HMPC controller fully resides in
the AM.
For the levels unit process optimization upto production optimization, still a lot of
software has to be written. For the calculations in these levels samples are taken of at
least one second. The hardware and availability criteria of these levels are not as
stringent as for the primary process control level. In these levels VAX computers and
PCs are typically used. Hardware and software breakthroughs can be applicated
relatively quickly. The quality of these levels is determined by software rather than
hardware. Appropriate algorithms are developed "bottom up" by process control
scientists and "top down" by management scientists.
48
7 Recent and future topics in Research
7.1 The Fundamental Control Problem [8]
As we have seen before a DCS with split architecture automates the four technologies
that are integrated in the decision making process: (i) measurement (ii) control (SISO
and MIMO layer) (iii) optimization (iv) logistics (scheduling and allocation of raw
materials). Such a control system extracts the most profit out of existing processes (no
redesign needed) while responding to market changes.
For still a lot of processes all over the world integrated control systems have to
be designed (15.000 to 20.000). Since each petro-chemical process is unique, costs
cannot be lowered by manufacturing at large scale. So design and maintenance costs
must be minimized. This requires a unified approach to every specific control
problem which starts from setting the specific problem into the framework of a
standard problem formulation: The Fundamental Control Problem. In this section a
design procedure to solve the FCP will be treated. The value of the FCP lies in the
fact that it limits the solution space to the solution(s) of the real control issue. The
steps in the off- or on-line design procedure will make clear in what directions future
research should take place.
The advantages of a unified approach to control is that design and mainten-
ance costs are decreased and that the focus of research is redirected from the
application of intuitive process models (training and experience) and the use of ad-
hoc solutions by non-expert designers to the fundamentals of the control problem
itself.
There are two major disadvantages to the unified approach: (i) The unified
approach forces one to model a process completely which is of course much more
difficult than modelling parts of the process driven by the conscientiously built up
library with standard solutions for standard problems. (ii) The treated unified
approach does not centre the real issue. It implicitly states that the quality of good
control is something that a process or model admits whilst the key issue that mainly
bothers academics is what control objective a certain model together with its uncer-
tainty description admit.
Before the design procedure will be treated the Fundamental Control Problem is
stated: On-line update the manipulated variables to satisfy multiple changing
performance criteria based on a process representation which includes a description
of the uncertainties. The practical performance criteria are:
economic:
safety and
environmental:
equipment:
product quality:
maintain PVs at targets dictated by the optimization phase
or dynamic minimal cost function
some PVs must not violate specific bounds for personal or
equipment safety or environmental regulations
physical limitations of equipment
consumer's specifications
49
human
preferences: operator will not tolerate certain jaggedness or oscilla-
tions, preferred modes of operation
These criteria have to be translated to mathematical expressions, to be divided in:
objectives: functions of variables to be optimized
constraints: functions of variables to be kept within bounds, two types:
hard constraints: no violations allowed
soft constraints: violations allowed temporarily for the satisfaction
of other criteria
The translation to objectives and constraints involves compromises and assumptions.
Artificial Intelligence facilitates the formulation of the practical criteria and the
translation to mathematical expressions for non-expert designers.
Mathematical models for petro-chemical processes are hard to formulate and
often contain uncertainty. Since uncertainties impose restrictions on the satisfaction of
the performance criteria, apart from the process model an explicit uncertainty
description must be available to perform the control system design.
In figure 7.1 the Fundamental Control Problem is given.
Objectives I
Control the Top Compo- min (Yi* - yjF, i = 1,2, '" nobj
sit ion at Its Target
m
Inequality Constraints I
Yilow ,; Yi ,; Yihigh
The Reactor Temp. Must
m j l D ~ m i s mjhigh
Not Exceed High Limit
ILimj I .. max move
Equality Constraints I
The Air to the Regen-
Yi = Yi
, erator Must Be on
mj =mj*
I
ITarget at All Times
Performance
Criteria
Process
Representation
Model Equations I
m
The Overhead Temp. Yi (k) = L: aijl' Lim; (k-f)
Drops Fast When
I' =,
+ di (k)
Steam is Reduced
>
Uncertainty Desuiption I
a;;, ' a;j' ,a;j,;" ~ I
Feed Composition Can
Change as Much as 20%
Iddk) I,; dj
in a Couple of Hours
Assumptions
Qualitative --- --- Quantitative
Compromise
Figure Zl The Fundamental Control Problem [8]
50
Objectives I Objectives I
Single
Lumped Multiobjeetive
Quadratic Quadratic
Setpoint Tracking Setpoint Tracking.
Input Penalties
Objectives I
Lumped Multiobjective
Quadratic
Setpoint Tracking.
Input Penalties
Objectives I
Lumped Multiobjective
Quadratic
Setpoint Tracking.
Input Penalties i""':":""""
I FiJCed
Objectives I
Lumped Multiobjective
Quadratic
Setpoint Tracking,lnput
r:C:;"""ha'--n-oii:-no-l
Changing
Inequality Constraints
Equality Constraints
Linear
MIMO
linear
MIMO
,.-------------.., r-------...,
I Equality Constraints
I
I
I
I
I
'-- LIC:.;.;h::.;.an;;;l.qi'c.;.nq....,
-
,
,
,
I
I
I l.. -'
- - - - - -
I , , II
, I , II
I I , II
I II II
I I , II
, II " .. ..J l.. ..... ...
... '- "" L. ..J
r-------------, r-------------.., r-------------.,
I Inequality Constraints I 'Inequality Constraints I 'Inequality Constraints
, , I , ,
I , , , I
I I I I'
, I I II
, I I I I
I , I II
Model Equations I
2nd Order
Linear
.5150
Model Equations I Model Eq uations I
Linear linear
MIMO MIMO
I ChanQinq
Model Equations I
Linear
MIMO
Model Equations I
linear
MIMO
r-------------, r-------------""\ r-------------.,
IUncertainty Description I ,uncertainty Description I 'Uncertainty Description I
I I I II I
,. Implicit It. Implicit I I- Implicit I
I. On-line Oetuning I I_ On-Line Oet:uninq I Ie On-Line Oetuning I
I , I 11 I
, , I II I
, I I II I L. ..J l.. ..1 ..J
ssv
Uncertainty Description
Explicit
.. Structured
Parameten and
Oi1turbances
, ,
,. Implicit I
I. On-Line Oetuning I
I ,
I I
I -,
STR 5150 P)[)
Figwe Z2 Control Methodologies in light of FCP [8J
In figure 7.2, 5 control methodologies are reviewed in the light of the FCP. The
elements that are not dealt with at all or implicitly only (comprising the elements in
boxes that are not dashed) are shown in dashed boxes.
In figure 7.2 (a) the model uncertainty can be dealt with by on-line detuning
making the setpoint tracking worse.
Structured Singular Value Synthesis (SSV) is a H inf technique. QDMC is a
model based multivariable control technique that minimizes a quadratic criterion on
line. On-line modification of all elements of the control structure is possible and
constraints can be added. QDMC belongs to the class of Model Predictive Control-
lers.
In 1988 the MPC was improved to include an uncertainty description which
provides a controller that deals with all elements of the FCP explicitly.
Now the complete design procedure will be described under the assumption that
there is enough computing capability and logical processing available to perform it
off-line which is by no means restrictive in the light of advances in the area of Super
Computing.
1 Provide the most accurate model and uncertainty description available for the
process. Requires a model building facility that combines white and black box
modelling techniques. This will demand advanced graphics facilities and
Artificial Intelligence technology superimposed on chemical engineering
technology.
2 Formulate the pratical performance criteria. Requires an advanced graphics
51
terminal interfacing an Expert System that contains Artificial Intelligence. The
AI links the graphics to practical criteria.
3 Translate the practical criteria to objectives and constraints (mathematical
criteria). Performed by an Expert System.
4 Solve the Fundamental Control Problem. If this optimization problem has no
solution then go back to Step 1 or 2 (or 3) to improve the process representa-
tion or relax the criteria. If the problem has a solution there is still no guaran-
tee that it can function on-line. If the hardware facilitates, the design is
complete. If not then proceed with Step 5.
5 Specify the controller equations (= expressions for the MVs). Pick a control
scheme that fits the available hardware. The equations that govern the chosen
control system are added as constraints to the FCP.
6 Solve the FCP with the controller equations substituted for the MVs. Now the
tuning parameters in the controller equations have become the independent
variables. If there is a solution the design is complete. If no solution then relax
criteria (Step 2 or 3) or choose different controller (Step 5).
In section 7.2 specific research for the solution of the FCP will be treated. In section
7.3 the place of the FCP in the design of Hierarchical Control Systems will be
outlined and in section 7.4 specific research for HCS will be described.
7.2 Areas of research for the solution of the Fundamental Control Problem [8]
Time domain optimization under uncertainties: The presence of time-domain inequality
constraints in the FCP forces the optimization to be performed in the time domain.
Techniques must be found that solve the optimal control problem in the face of
uncertainties in the time domain. (See figure 7.2: QDMC deals with model uncer-
tainties implicitly and SSV does not deal with constraints). Existing techniques (1
1
-
optimal controller, newly developed) do not perform well. This, together with the fact
that models are not accurate enough makes the constraint handling still takes place in
nonlinear logic elements with exactly specified priorities.
Multi-objective optimization techniques: All control methodologies of figure 7.2 handle
the satisfaction of multiple objectives by weighting them into one. There is a need to
develop optimization techniques that consider multiple objectives and allow a
transparant specification of priorities. Expert Systems might help.
Nonlinear process representations: Improvements on the performance of control
systems can be made by utilizing nonlinear models in for example adaptive control
structures. Generic nonlinear models for common processes and methods to validate
their parameters with on-line experiments need to be developed. Again Expert
systems can play a role in offering these models to control engineers.
Uncertainty descriptions: The techniques available to describe uncertainties rely on
linear models. This leaves us with a lot of problems related to unstructured represen-
tations because of nonlinearities. The use of nonlinear models would provide
parameters with more physical significance, though taking more hardware capability.
52
The designer should be provided with tools to convert from a nonlinear to a linear
process representation with uncertainties.
Artificial Intelligence [16]: AI research provided rule-based systems that describe
behaviour.
Step 1, 2 and 3 of the design procedure of section 7.1 require Expert Systems
containing Artificial Intelligence. In general these systems will help in facilitating the
synergism between qualitative and quantitative knowledge bases and requires the
least interpretation by the designer. AI technologies must help closing the gap
between what we can realistically implement on-line and what we can solve off-line in
order to provide the non-expert user of the control system with any level of the core-
control theory.
The development of Expert Systems requires new computing environments and
programming styles whose central features are; object-oriented programming and data-
or result-driven procedural programming [16].
Super Computing: On the hardware side, advances in the field of Super Computing
indicate that there is the capability to execute the kind of optimization problem in the
design procedure, not only off-line but on-line as well.
However, large difficulties are encountered on the algorithm and software side.
The FCP for a system with 5 inputs, 7 outputs and 70 states cannot be solved with
CONSYD, HONEY-X and MATLAB [13]. Model reductions were not possible
either.
7.3 The Fundamental Control Problem versus the Hierarchical Control Structure
In section 7.1 the FCP was described. The design procedure arrives at an optimal and
centralized solution that could be an optimal robust or model predictive controller.
However these are not forms that can be readily implemented.
In practice, a hierarchical control structure is preferred for reasons mentioned
in chapter 3. However, this practical design procedure focusses away from the FCP. It
starts with the selection of the variables which are to be manipulated and measured
which already determines the performance of the control system. Then the controller
structure is chosen: which MVs are changed and which errors determine these chan-
ges. This procedure makes it possible to incorporate reliability which is something
industry appreciates more than "optimality frequently on manual".
To improve the "practical" design procedure, techniques have been developed
to translate overall performance and robustness specifications into specifications on
individual loops [13]. Designing the individual loops according to these specifications
guarantees the satisfactory performance of the overall system. The major deficiency of
these systems is their conservativeness that can exclude solutions that meet the
specifications.
One can conclude this section by saying that the Fundamental Control
Problem or finding the optimal control solution, has not the first priority for the
design of Hierarchical Control Systems. This is the reason why safety and operability
provisions can be found in the lower levels with small time constants and optimality is
the responsibility of the highest levels with large sample periods.
53
7.4 Areas of research for the Hierarchical Control Structure
Attempts to systemize the practical control structure problem provided the following:
There is little need for a systematic procedure [13] to fill in the primary
control level (figure 3.2) of the hierarchical control structure. The choice of feedback
and feedforward loops is usually straightforward and can be handled by engineers
with limited experience.
The unit process control or supervisory level is problematic. The model that
resides in this level is often not accurate enough to protect against constraint
violation. Research circles neglected the design of practical features like anti-reset
windup schemes, selectors etc. for this (multivariable) level [13].
As stated in section 3.2.1 the dynamic plant performance optimization level is
based on steady state plant simulation which malfunctions because steady states are
not always reached during production nowadays. Instead rigorous dynamic models
have to be developed to perform the simulation.
7.5 The role of Academia in Control Research [14]
In "measurement and control" there are two types of researchers: academic and
industrial. Since there is no reason to compete, both parties should cooperate. The
role of academia in control research could consist of:
develop and prove control concepts since industry is unwilling to upset current
production procedures
develop the basic theory behind control algorithms and procedures whether
developed in industry or academia
focus more than industry does (has a more immediate focus) toward high risk
but potential large pay-off problems
innovation and new ideas
Aid of industry for academic research could consist of:
direct assistance and support
communication: despite painful for the company involved it is helpful for an
academic to know what does not work in the real world; industry should
promote their failures
54
8 Process Automation Projects
Automation projects in process industry significantly differ from administrative and
general (computer system and network software) automation projects nowadays.
Some differences will be described here; General automation projects are
often performed by software houses or by specialists that are in detachment. Process
automation projects are carried out on a fixed price basis whilst competition forces
the prices to be low. Excessions are generally paid by the control system supplier.
Administrative and general automation projects often deal with rapidly changing
central computer systems. This reduces differences between consequent projects and
eases maintenance and extensions.
In process automation projects at least three parties are involved; the principal
(site), contractor (engineering consultants / "knowledge tank") and the control system
supplier. The site brings in people with expert knowledge about the process to be
automated, the contractor subdivides the job in parts and "contracts" suppliers of
specific hardware and sometimes software. The contractor consists of a team of
measurement and control technologists. The control system supplier employs technol-
ogists dedicated to the control system.
Process automation projects know exponentionally raising costs towards the
end of the project. This is caused by a differentiation of knowledge over the parties
involved rather than poor technology (proven before) which results in difficulties in
understanding each other, elaborating on minor issues whilst the control strategies
become ever more complex. The majority of the problems in process automation
projects originate in the Function Design Specification (FDS) phase. It appears to be
very difficult to specify the funtional requirements in an unambigious way such that
technical design, implementation and test proceed smoothly. FDSers (site) very often
have a chemical origin with little process control knowledge. Before the suggestion of
possible solutions to this problem, the main priorities of all parties in one project are
highlighted:
site: safety, keeping process in operation, optimal production
contractor: good cooperation between all parties, no delays
control system
supplier: ultimate costs including necessary benefits match fixed (a priori)
pnce
As was proven in the chapters before, the "safety" and "operation" priorities of the site
are guaranteed nowadays due to the experience that control system designers gained
over time. Since these conditions are fulfilled sites are more interested in optimal
production, in other words in the relation between the Fundamental Control Problem
and the Hierarchical Control Structure. Research!
The first priority of the contractor is a good cooperation due to a good
coordination in order to gain the confidence of the site and be sure of orders in
future time. Delays are very precious.
55
As mentioned before, due to the fixed price principle, exceeding costs are
covered by the control system designer himself. Because "the hours spent" are the
biggest threat to profits, control system designers ought to be interested in Computer
Aided Design facilities. AI research!
Proper education and research could solve the problems. It would be sensible
to raise post-graduate research schools that teach students the system engineering
approach [15]. It ought to be a post-graduate school since system engineering is
founded on generalized, sometimes abstract and theoretical concepts. Such a school
could have a curriculum for chemical and measurement-and-control engineers,
spending their first year(s) by extending their control- and chemical / physical
knowledge respectively. Such schools could deliver all-round technologists with
"helicopter views" on chemical processes and measurement-and-control concepts
(good FDSs) to make the process automation projects a success for all parties.
The research activities of these schools would cover how Hierarchical Control
Systems are able to control as good as possible (the relation between the Fundamen-
tal Control Problem and the Hierarchical Control Structure) for high risk applications
with a potential large pay-off and the automation of the control-system-design tasks.
A trusted representative of these schools could gain information about failures in
process industry that must remain secret.
The post-graduated students would transfer the research knowledge to
industry.
56
9 Conclusions
The structure of the IDC 3000 control system of Honeywell conforms the ideal
hierarchical control system structure. A MIMO Internal Model based Controller was
implemented succesfully which proves that the unit process control can be executed
optimally in IDC 3000. There will be more than enough room to implement an
additional constraint handler (which will demand high quality of model) and essential
features like anti-reset wind up schemes.
The identification and controller design part of the Model Based Control
package HMPC is not DCS resident which delays serious use. It can be improved
through application of algorithms that arrive at more compact models, decouple the
control and optimization problem and control "robustly". Combined with the Statisti-
cal Process Quality Control, a Model Based Controller can be powerful to processes
that have fast and slow (disturbance) dynamics.
The levels unit process optimization, dynamic plant performance optimization
and production optimization needn't be DCS resident. The slow sample periods for
these levels relief hardware criteria. The dynamic plant performance optimization
level requires the development of rigorous dynamic models. The development of
software for these levels will continue.
The Fundamental Control Problem formulation, suggested by Shell, decreases
design and maintenance costs but cannot be combined with the design of DCSs which
inherently focusses away from the FCP. The FCP causes specific research of which
the application of Artificial Intelligence in Expert Systems to facilitate CAD is
perhaps the most interesting for DCSs. Nowadays the proper CAD environment does
not exist to solve the FCP for advanced controllers.
The increasing complexity of process automation projects asks for a system
engineering approach to be taught to graduates at research schools that will examine
high risk but potential large pay-off problems. Industry should assist, support and
communicate.
57
References
[1] J.C.J. van de Wiel (1991)
Het traject aanvraag tot en met overdracht bij hedendaagse industriele
procesautomatiseringsprojecten (ATEMO)
Uitgave afdeling Industriele Automatisering, Honeywell
Amsterdam
[2] A.AH. Damen en H.H. v.d. Yen (1990)
Dictaat van het college Moderne Regeltechniek
Faculteit Elektrotechniek, Vakgroep Meten en Regelen
[3] J. Borer (1985)
Instrumentation and Control for the Process Industries
Elsevier Applied Science Publishers, London and New York
[4] A.C.P.M. Backx (1993)
Industrial Application of Process Identification Techniques
Proceedings Symposium "Procesmodellen in het regelalgoritme"
June 3 1993, Eindhoven University of Technology
[5] C.W. Hart and A.E. Ogden-Swift (1993)
Industrial Application of Model Based Predictive Control
Proceedings Symposium "Procesmodellen in het regelalgoritme"
[6] Honeywell bookset on mc 3000
[7] W. Giloi (1966)
Stand und Aussichten der Rechneranwendung in industriellen Prozessen
Fortschritt Berichte VOl Zeitschrift, Dusseldorf Reihe 8 Nr. 7
[8] C.E. Garcia and D.M. Prett (1986)
Design Methodology based on the Fundamental Control Problem Formulation
The Shell Process Control Workshop, December 15-18, 1986
Butterworths, Stoneham, MA
[9] M. Morari (1986)
Three Critiques of Process Control revisited a decade later
The Shell Process Control Workshop, December 15-18, 1986
Butterworths, Stoneham, MA
[10] The second Shell Process Control Workshop 1988
Butterworths, Stoneham, MA
58
[l1J A. v.d. Draaij
Documentation on inplementation of IPCOS MIMO Controller in IDC 3000
Honeywell Amsterdam, Division Industrial Automation
[12J "De personal computer gids"
[13J M. Morari (1986)
Process Control Theory: Reflections on the Past and goals for the next Decade
The second Shell Process Control Workshop 1988
[14J B.R. Holt (1986)
Control of Autoclave Processing of Polymeric Composites
The Shell Process Control Workshop, 1986
[15J O.A. Asbjomsen (1986)
A Systems Engineering Approach to Process Modelling
The Shell Process Control Workshop, 1986
[16J D.B. Garrisson and D.M. Prett (1986)
Expert Systems in Process Control and Optimization: A Perspective
The Shell Process Control Workshop, 1986
59