SIMATIC PCS 7 APC-Portfolio

White Paper
How to Improve the Performance of your Plant Using the
Appropriate Tools of SIMATIC PCS 7 APC-Portfolio?

October 2008

For many customers (facility operators, engineering companies,

EPCs, system integrators) it is good news, that APC (advanced
process control) functions are increasingly offered as embedded
functions in distributed control systems, like for example by Sie-
mens in the SIMATIC PCS 7 APC-Library and the Advanced Process
Library. Looking at the large variety of different APC tools, there
is a choice: which approach is appropriate for which type of task
or process, in order to improve plant performance and achieve
the maximum benefit with the minimal effort.
The following contribution starts with a survey of popular APC
tools as offered in the PCS 7 APC-Library. Based on this, a typical
procedure to improve process control using APC tools is outlined
and illustrated by a case study.

A white paper issued by: Siemens. © Siemens AG 2008. All rights reserved.
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 2

Contents

1 Survey: APC Tools in the PCS 7 Libraries ................................................................... 3

1.1 Introduction .......................................................................................................... 3
1.2 Control Performance Management ....................................................................... 3
1.2.1 Control Performance Monitoring ..................................................................... 3
1.2.2 Computer-aided Controller Optimization (PID-Tuning).......................................5
1.3 Extensions to PID Control...................................................................................... 5
1.3.1 Override Control ............................................................................................5
1.3.2 PID Gain-Scheduling .......................................................................................6
1.3.3 Smith Predictor Control for Deadtime Processes................................................ 7
1.3.4 Dynamic Disturbance Compensation (Lead-Lag Feedforward Control) ................ 8
1.4 Multi-Variable Control (MPC: Model Predictive Control) ..................................... 10
1.4.1 Intuitive Explanation of Predictive Control...................................................... 11
1.4.2 DCS embedded MPC versus MPC with Online Optimization ............................ 12
1.4.3 Configuration of Predictive Controllers .......................................................... 13
2 Typical Procedure to Improve Plant Performance using APC Tools.......................... 14
2.1 Situation Analysis and Potential ......................................................................... 15
2.1.1 Identification of Economic Targets for Process Control .................................... 15
2.1.2 Analysis of Current Plant Status incl. Control Performance Monitoring ............. 16
2.1.3 Review of Basic Automation, Eventually PID-Tuning ........................................ 16
2.2 Concept Definition of APC-Functions .................................................................. 16
2.2.1 Selection of Appropriate Approach Based on Analysis ..................................... 16
2.2.2 Design of Control System Structure, Selection of MVs, CVs and Constraints ...... 18
2.3 Configuration of APC-Functions .......................................................................... 19
2.3.1 Process Excitation and Recording of Learning Data ......................................... 19
2.3.2 Modelling .................................................................................................... 19
2.3.3 Controller Tuning ......................................................................................... 20
2.3.4 Commissioning ............................................................................................ 20
2.4 The Most Important Rules at a Glance ................................................................ 21
2.5 Documentation, Training and Maintenance ........................................................ 21
3 Case Study Distillation Column............................................................................... 21
3.1 Template-based Implementation ........................................................................ 22
3.2 Economic Benefit ................................................................................................ 23
3.3 Summary............................................................................................................. 24
4 Literature................................................................................................................ 24

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 3

1 Survey: APC Tools in the PCS 7 Libraries

With the APC-Library of Simatic PCS 7 ( [1.], [2.]),

for the first time advanced process control func-
1.2 Control Performance Ma-
tions are included in the scope of delivery of the nagement
distributed control system (DCS). Besides the core
control algorithms, there are also the related proc- The generic term control performance management
ess tag types and software tools for computer-aided includes control performance monitoring and opti-
controller configuration available without extra mization of control loops.
charge. 1.2.1 Control Performance Monitoring
The APC functions can be divided in three classes: Empirical studies have shown that many of the
• control performance management, control loops in process industries do not fulfil their
requirements properly and there is large potential
• extension to PID control, and for improvements. However not everybody is aware
• multivariable control. of this.

Fuzzy Control and artificial neural networks are

also offered in the context of Simatic PCS 7 in the open
loop
excellent

so called AddOn Catalogue [3.], but are not further only 1/3 of the plant
loops are running in
detailed in this paper. 16% good performance !

36%
Due to the reduced costs of a system-embedded
and template-based implementation, a large num- 16% acceptable

ber of applications including small and medium

sized processes become accessible for advanced 2/3 of the plant loops
process control – including processes where it is 10% 22% have room for
improvements or have
not affordable (in the context of desired amortiza- urgent need for action !

tion times) to interface expensive external ad- poor fair

vanced process control software packages to the Source: Control Engineering May 2008

DCS.

Figure 1: Potential for improvements in control performance,

1.1 Introduction from [14.]

In the following sections, each of the APC tools is Plant operators or instrumentation and control
explained by its main principle, typical use cases technicians on their own do not have a chance to
and typical applications (industry branches, process permanently supervise the huge numbers of control
units). The short descriptions are intended as intro- loops they are responsible for. Therefore functional-
ductory reading and make no claim to be complete. ity is needed to automatically monitor the control
Further information can be found in the literature performance of all loops in a plant, and all the time,
cited and the relevant product documentation. in order to schedule specific maintenance activities
or selective controller re-tuning in a timely manner
The PCS 7 example project ‘Getting Started with
if the performance of single control loops is de-
Advanced Control’ allows users to get familiar with
creasing or troubles are developing. For monitoring
the new advanced process control structures by
in the sense of a non-invasive diagnostic, only the
doing hands-on experiments without having to
measurement data of regular process operation are
intervene in the real process. This way, the concept,
evaluated. (Controller re-tuning however requires
the requirements and the benefit of a certain APC
active plant experiments.)
structure can be realized before it is applied to a
real process. For this the examples include process Each control loop to be monitored is by default
simulation and intuitive OS-pictures. For the exten- equipped with a dedicated performance monitoring
sions to PID control, the same process simulation is function block (ConPerMon, [6.] ), like designated
set up twice, one instance with APC extension and in the process tag types of the Advanced Process
the other without - all other process and controller Library [9.]. Using these process tag types, the
parameters identical. The advantages of the APC engineering effort for manual linking of monitoring
extension can be tested as a direct comparison. and controller function blocks can be reduced.

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 4

• A mean steady-state control deviation ≠ 0 at a

Control Performance Index

constant setpoint is an indication of problems in
the control loop if the controller has integral ac-
tion. You should then check the following poten-
ConPerMon tial causes:
• (1) The actuator does not have sufficient capac-
ity. As a result, the controller's manipulated sig-
nal constantly approaches its limit. This can be
SP PID Process PV caused by unsuitably dimensioned actuators,
Controller
+ - varying operating conditions or simply by wear
and tear.
• (2) The manipulated variable demanded by the
controller does not have an impact on the proc-
Figure 2: Signal flow for control performance monitoring, SP:
ess, for example because the actuator is defec-
Setpoint, PV: Process Value, MV: Manipulated Variable
tive.
• If the process gain (e.g. the intensity of heat or
In steady state process operation this monitoring material exchange) changes gradually as time
function block calculates the following stochastic progresses, this is an indication of wear phe-
features of control performance: nomena in the process, such as fouling of heat
• Mean value, variance and standard deviation of exchangers, valves or shutters, decreasing effi-
controlled variable, ciency of process units, etc. If, for example, a
temperature is regulated by a heat exchanger
• Mean value of the manipulated variable and
and fouling is developing on the exchanger sur-
control deviation, faces, the heat transfer coefficient, and conse-
• Control performance index, quently the process gain, is reduced. Within cer-
tain limits, this can be compensated for by a
• Estimated steady state process gain.
closed control loop (so that the controller ini-
For setpoint steps, the following deterministic fea- tially hides away the problem). Although the
tures of control performance are evaluated: original control loop dynamics can be restored
• Rise time, settling time and settling ratio, (to a certain extent) by suitable increase of the
controller gain as the fouling increases, it is ad-
• Overshoot absolute and relative to the step
visable to eliminate the cause of the problem; in
height. other words, to clean the heat exchanger.
Other statistical and graphic evaluations of the • The control performance index (CPI) in the unit
signals in the control loop over longer, freely se- [%] describes the current variance of the con-
lectable periods are available in the faceplate of the trolled variable relative to a reference variance
ConPerMon block.
σ ref 2
In an overview representation of a plant or unit, (benchmark). It is defined as ξ= 100% .
you can obtain a clear picture of the status of all σ y2
control loops using ConPerMon block icons (indica-
tor light function). The aim is to detect problems as • The CPI moves in the 0 < ς ≤ 100% range. If the
they develop and to focus the attention of the user current variance corresponds to the reference
on the control loops in a plant that are no longer variance, the index reaches the value 100. Ide-
operating correctly. ally, the reference variance is obtained in a de-
fined good state of the control loop (e.g. after
Application Examples commissioning with a PID-Tuner) and stored
Control performance monitoring can be applied in when the ConPerMon block is initialized. If the
any process and any industry. It provides a fast current variance increases, the control perform-
overview of problem loops and their deviation from ance index drops accordingly. If the performance
ideal performance inside the DCS. Typical applica- index falls below a specified threshold, a mes-
tion areas are plants with huge numbers of control sage is generated, and the process picture sym-
loops, as in the steel or paper industry, refineries, bol shows a colour change from green to yellow
petrochemicals and bulk chemicals. Several conclu- or red. Decreasing control performance means
sions with respect to the state of the corresponding that controller tuning and process behaviour do
loop can be drawn by interpretation of the indica- not fit together any more. If the initial process
tors calculated by the monitoring block, without behaviour cannot be restored, it is advisable to
special prior knowledge: re-tune the controller.

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 5

• The relative overshoot as a percentage is a kS

measure of the damping of the control loop. If g ( s) =
(t1s + 1) n
relative overshoot is more than 20 or 30%, the
loop gain (gain of the controller multiplied by Only the three parameters process gain k S , time
the gain of the controlled process) is generally
constant t1 and system order n have to be esti-
too high, either because the controller was badly
tuned from the beginning or because the proper- mated by the PID-Tuner. The bigger the order n ,
ties of the controlled process have changed over the bigger is the time lag relative to the settling
the course of time. If overshoot is significantly time of the step response.
too high, the control loop is generating weakly The calculation of the optimal controller parameters
damped oscillations in the plant. If overshoot is according to the modulus optimum [5.] is based on
too high, it is often helpful to reduce the gain of the identified process model.
the controller.
Moreover, the PCS7 PID-Tuner offers the possibility
1.2.2 Computer-aided Controller Opti- to choose between two different variants of con-
mization (PID-Tuning) troller design:
Many PID-controllers in industry are tuned by trial- • Optimal disturbance compensation. (The trade-
and-error methods or by heuristic rules, and the off are 10-20% overshoot for a setpoint step.)
differential action is frequently not considered at
• Optimal setpoint tracking without overshoot.
all. For certain standard control loops like the flow
This can be achieved by structure decomposi-
control of fluids with a proportional valve, there are
tion (P/D-action moved to feedback path) with-
empirical values for standard parameter sets. For
out losing in disturbance compensation only
slow controlled processes like temperature control
for low order plants, whereas controlled proc-
loops, an optimization by trial-and-error takes too
esses with order bigger than 2 require a gain
much time, because the observation of a single
reduction.
step response may need several hours.
1.3 Extensions to PID Control
This generic term summarizes different solution
PID Tuner approaches that can be realized by clever combina-
tion of PID controllers with other standard function
Controller blocks, and that are offered as process tag types in
Parameters
PCS 7 Advanced Process Library [9.].
PID Process
SP
Controller
PV
In this area, there is no generally-accepted distinc-
+ - MV
tion, as to which of these structures are to be called
“advanced” control. The following process tag types
from [1.] are considered “conventional” control in
Figure 3: PID-Tuning (Computer aided control system design)
this paper and therefore not further described here:
cascade control, ratio control and split-range con-
trol.
Consequently the application of computer-aided
controller design tools is winning recognition. The 1.3.1 Override Control
calculation of optimal controller parameters is per-
In override control, two or more controllers share a
formed with an experimental procedure starting
common actuator. Depending on the current proc-
with the modelling of the process dynamics. The
ess state, a decision is made as to which controller
process is excited with a step of the manipulated
actually has access to the actuator; in other words,
variable or a setpoint step (if there is at least a sta-
the various controllers can override each other.
ble but suboptimal controller setting). A dynamic
process model is estimated from the stored meas- A typical use case is a gas pipeline with pressure
urement data by the tuning tool, i.e. the process and flow control using a single valve. The main aim
parameters are calculated such that the learning of the control is to achieve a certain flow rate,
data are fitted optimally (in a least squares sense) however due to safety considerations, the pressure
by the model. must be kept within certain limits. The pressure
controller is therefore known as the "limiting con-
In the approach of the PCS 7 PID-Tuner [8.], which
troller" or "secondary controller"
is especially simple and robust and fully integrated
into the DCS engineering software, an algorithm
with PTn-models of rising order is applied:

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 6

its of all controllers of the currently highest

Decision
Logic (lowest) manipulated variable must be cor-
rected slightly up or down by, for example, 2%
of the manipulated variable range. This means
SP1 - PID control 1
(z.B. FIC) that this scheme can also be used in applica-
+
tions with more than two controlled variables.
Process
There is no Windup problem at the high limit
because the highest manipulated variable takes
PID control 2
SP2
(z.B. PIC) over control anyway. This approach avoids the
+ -
limit cycle oscillation of alternative 1 but is, in
principle, asymmetrical. In other words, either
a high or a low limit of the secondary con-
Figure 4: Override control
trolled variable can be monitored but not both.
This type of override control is described in
most control textbooks, particularly in the USA.
The logical decision as to which controller should
It can, however, only be used with PID algo-
be active can be made based on two different crite-
rithms that allow online manipulation of the
ria resulting in two different types of override con-
MV limits (manipulated variable limits, in PCS 7
trols:
as of V6.0). Alternatively, in PID controllers
• The decision is based on a measurable process based on an incremental PID algorithm, there is
output variable, for example one of the two the possibility to overwrite the MV value of the
controlled variables. In the example above, the last sampling step (that is stored inside the
warning limits of the pressure controller can be controller) by a linkable input variable called
used to decide whether the pressure controller “external reset”. Incremental PID algorithms are
should be active. The passive controller is in typically offered by US American DCS vendors
tracking mode to avoid Windup problems and and in PCS 7 APL as a special function block
to ensure bumpless transfer. The setpoint of PIDConR [2.].
the secondary controller must be somewhat
Further Application Examples
lower than the switchover threshold so that the
• Steam generator: The primary controlled vari-
transfer can be reversed again. This type of
able is the steam pressure but the water level
override control is easy to understand and to
in the steam tank must be monitored so that
implement. Its advantage is that the high and the heating coils remain completely covered by
low limit of the secondary controlled variable water and the tank does not overflow. The only
(for example pressure) can be monitored; its manipulated variable is the outlet valve.
disadvantage is that a limit cycle oscillation re-
sults as soon as the limiting controller needs to • Compressor: The primary controlled variable is
intervene. The secondary controller will always the throughput but the pressure must be moni-
tored to make sure it does not exceed a safety
attempt to return its controlled variable to the
limit. The only manipulated variable is the mo-
safe range and to return command to the main
tor speed.
controller (for example flow rate) so that the
active and passive controllers swap over con- • Steam distribution system: Every plant involv-
tinuously. This variant is therefore only recom- ing industrial processes has a network of pipes
mended when the secondary controller is sel- to distribute steam at various pressures
dom required and functions mainly as a safety throughout the plant. The high pressure of the
or backup system. steam is reduced to lower levels via a valve.
The primary controlled variable is the pressure
• The decision is based on a comparison of the at the lower-level stage, however the pressure
manipulated variables of both controllers, for in the high pressure piping must also be moni-
example the controller that demands the tored to make sure that it does not exceed a sa-
higher (or lower) controlled variable takes con- fety limit.
trol of the actuator. In the example above, the
controller that wants to open the valve further
1.3.2 PID Gain-Scheduling
takes over control. The setpoint of the secon- Many processes have a non-linear response due to
dary controller defines the switching threshold. non-linear physical, chemical or thermodynamic
Both controllers run the entire time in auto- effects. When such a process needs to be kept in
matic mode. To avoid Windup problems, the the close vicinity of a fixed operating point, the
manipulated variable limits must be tracked in transfer response can be linearized around this
a crossover structure: When the higher (lower) operating point. A linear PID controller can be de-
manipulated variable wins, the low (high) lim- signed for this linearized transfer function. If, how-

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 7

ever, the process has a strongly non-linear response view. This becomes clear even with benign non-
and/or operates at different operating points, no linearities (that are continuous and can be differen-
constantly good control response can be expected tiated) when setpoint step changes are made be-
throughout the entire operating range. Due to the tween different operating points. With non-
non-linearity, various gain factors or process time linearities that are discontinuous or cannot be dif-
constants are in effect at different operating points. ferentiated or with non-monotonic non-linearities,
In keeping with this, different controller parameters great caution is needed.
will be considered to be optimum

Measurement (X)
Application Examples
• Control (especially temperature control) of
batch processes, for example, batch reactors
Gain and batch columns
Scheduler
• pH value control by dosing acid or base
Control (nonlinear titration curve)
parameter sets

SP PID Process PV
• Temperature control with phase transitions (for
Controller example, fluid/vapour)
+ - MV
• Control of semi-batch plants (continuous plants
with operating point changes, for example, po-
lymerization reactors)
Figure 5: Gain scheduling
• Control in power plants with load changes

1.3.3 Smith Predictor Control for Dead-

One possible (the simplest) solution to this problem time Processes
is known as gain scheduling or parameter schedul-
ing. Using a tool such as the PCS 7 PID Tuner, vari- A deadtime can be recognized from the observa-
ous experiments are performed at different operat- tion, that after intrusion of an MV move, there is no
ing points, in each case with low signal amplitudes. reaction of the controlled variable at all for a cer-
This results in different PID parameter sets for each tain time (the deadtime). In processes with large
operating point. Up to three such parameter sets dead times θ (e.g. θ > 0.25 t1 relative to the domi-
can be stored in the gain scheduling function block nating lag time constant t1), a standard PI controller
(GainSched). The suitable parameter set is selected must be tuned very slowly and compromises must
depending on a continuously measurable variable therefore be accepted in the control performance.
(measurement X in Figure 5) that describes the state The control performance can be significantly im-
of the process, typically the control variable PV proved with a Smith predictor that can be derived
itself. Between the operating points for which there from the IMC principle (Internal Model Control) of
are exact parameter values, the values are calcu- model-based control.
lated by linear interpolation of the neighbouring
MV
interpolation points so that soft and bumpless tran- SP PID g(s)
Process
PV
e-θs
sitions are possible between the operating points. + - + -
Controller

The term "parameter scheduling" makes it clear that Deadtime

the "timetable" for adjusting the parameters is Model
e-θs +
gm(s)
specified in advance. In contrast, an adaptive con- -

troller adapts itself automatically to the differing

process response during operation.
The function block GainSched is produced from the Figure 6: Smith predictor
CFC chart "fbGainSched" by compiling it as a block
type. This CFC chart is supplied with the library [1.]
so that the user has the option of expanding the
To achieve this, the transfer function
existing basic functionality as necessary, for exam-
ple to more than three operating points, or applica- g s (s ) = g (s )e −θs of the controlled system is split
tion specific logic for selection of parameter sets. up into a part g(s) without dead time and a purely
−θs
Note: The combination of several locally optimized dead time part e with dead time θ. Only the
controllers by gain scheduling to form a non-linear controlled variable (PV) affected by dead time can
controller does not necessarily represent an opti- be measured in the real process. However, a virtual
mum non-linear controller for the non-linear proc- estimate of the controlled variable free of dead
ess when considered from a mathematical point of time can be taken from the process model (that will

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 8

become part of the controller) and fed to the con-

Disturbance
troller. This means that the controller itself can be DV
transfer
designed for the process without a dead time and function gz(s)

can therefore be tuned much more tightly. To com- Compensation

c(s)
pensate for unknown disturbances, an estimate of
the controlled variable affected by dead time is
made in the model and compared with the genuine SP PID + Process + PV
Controller g(s)
measured controlled variable. This difference is also + - + +
MV
fed back to the controller.
In terms of practical application, it must be pointed Figure 7: Feedforward disturbance compensation, DV: Distur-
out that the performance of the Smith predictor
bance Variable
depends largely on the model fit; in other words,
the dead time must be known. The dead time must
be constant or its value must be permanently
adapted. The impact of a measurable disturbance can be
estimated in the form of a transfer function
y (s )
Note: For processes with large dead times, a model
predictive controller (c.f. section 1.4) is also suit- g z (s) = (with y: CV and z: DV) when the
able in a single-input single-output situation. It z (s )
provides greater flexibility in system modelling and controller is running in manual mode so that no
is more convenient thanks to the integrated design changes whatsoever to the controlled variable y =
tool. However, it does require more CPU resources, PV are caused by the manipulated variable and all
and does not allow online adaptation of the dead- changes can be attributed to the disturbance z(s).
time value.
The transfer function of an ideal feedforward con-
Application Examples trol c(s) can be derived from the requirement that
The typical cause for deadtimes in process engi- the impact of z on the controlled variable y
neering plants are run durations of fluids or gases should be zero for any disturbance signal z(s)
in pipes, or run durations of bulk solids on conveyor !
belts. g z ( s ) z − c( s ) g ( s ) z = (g z ( s ) − c ( s ) g ( s ) ) z = 0
• Temperature control via feeding of hot steam
or cold/warm water in a chemical reactor To meet this equation, the compensation block
jacket. After opening the valve, it takes some must approximate the equation
time until the temperated medium reaches the
jacket via the pipe. g z ( s)
c( s) =
g (s)
• Temperature control in chemical reactors or
distillation columns via external heat exchang-
as well as possible. For this to happen, the distur-
ers. After an MV move at the heat exchanger it
takes some time until the temperated medium bance transfer function must be known and the
flows back from the heat exchanger to the re- transfer function of the main controlled system
y (s )
g (s ) =
actor or column via the pipe.
(with u: MV) must be inverted.
• Load control on a conveyor belt: the spatial u (s )
distance between MV intervention and meas-
urement system can be converted directly into If both transfer functions can be modelled as first
a deadtime via the conveyor speed. k S −sθ
order plus dead time g ( s ) = e and
1 + t1s
1.3.4 Dynamic Disturbance Compensa-
tion (Lead-Lag Feedforward Con- kS z
g z (s) = e − sθ z , and θ < θ z applies, the result-
trol) 1 + t1z s
Feedforward disturbance control can be used when ing compensation element must represent the lead-
a known, strong disturbance affects the process lag transfer function
and its cause can be measured. In these cases, the k S z 1 + t1s − s (θ −θ ) 1 + t d s − sθ c
following general strategy applies: "Feedforward c( s) = e z = kc e
k S 1 + t1 z s 1+ t s
control as much as possible (as much as known in
advance and described by a model), feedback con-
This transfer function can be created outside the
trol as much as necessary (the rest including the
controller with a combination of elementary func-
model error and immeasurable disturbances)".
tion blocks.

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 9

An additional input at the PID controller block al- perature of the medium are the measurable
lows adding this signal to the MV value. It is impor- disturbance variables.
tant that the addition of sideline contributions to • Fill level control in a drum steam generator
the MV value is performed before the MV limitation using the inlet volume: the outflow is the
of the controller block, in order to ensure proper measurable disturbance variable that is deter-
limitation of the overall MV (including anti- mined by the variable steam consumption in
windup). the plant.
However for general transfer functions g (s ) and • Temperature control in a distillation column
g z (s ) there will be more complicated or even un- using the reflux ratio or heating steam flow:
the measurable disturbance variable is the mix-
feasible compensation functions. Those have to be
ture feed flow.
simplified by order reduction, which might reduce
the efficiency of disturbance compensation. This • Temperature and concentration control in an
simplification can go that far, that the process dy- agitated tank reactor using cooling medium
namics is completely neglected, and only flow and discharge volume: the temperature
and possibly also the concentration of the in-
kS z flow are measurable disturbance variables.
c( s ) = is implemented (static feedforward
kS Note: Dynamic disturbance compensation can also
control). be realized with a model predictive controller (c.f.
section 1.4), in multi-input multi-output and in
Application Examples
single-output constellations. It provides greater
• Temperature control of an industrial oven: at
the oven inlet, the disturbance variable feed flexibility in system modelling and is more conven-
flow is measured and fed-forward to the output ient thanks to the integrated design tool. However,
of the temperature controller. The impact of it does require more CPU resources.
varying flows on the oven temperature is an-
ticipated and compensated for by modifying
the heating power.
• Controlling the outlet temperature of a heat
exchanger via steam pressure or heat-
ing/cooling medium flow: flow and inlet tem-

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 10

Figure 8: Benchmarking template for disturbance compensation, taken from example project of PCS7 APC Library. Black: set-
point, green: controlled variables, orange: manipulated variables, light colours: with disturbance compensation, dark colours:
without compensation, purple: disturbance variable.

1.4 Multi-Variable Control

G(1,d)

(MPC: Model Predictive G(1,1) + CV1

+ +

Control) DV1
ModPreCon MV1 G(2,1)

G(1,2)
If there are several manipulated and controlled MV2
variables (MVs and CVs) in one unit, that are G(2,2) +
interacting with each other, you are dealing with + + CV2

a multivariable control situation (MIMO control: G(2,d)

multi-input multi-output, as opposed to SISO

control: single-input single-output). The impact
of each MV to each CV is described by a part Figure 9: Multivariable control. CV: Controlled Variable,
transfer function. MV: Manipulated Variable, DV: Disturbance Variable. Each
part transfer function G(i,j) describes the impact of MV j
to CV i. The figure shows a 2x2 process, but there can be
more MVs and CVs.

The aim of any control system is to bring each

controlled variable to its individual setpoint,
independent of the other CVs. This is difficult
because a move of one MV (e.g. MV1) does not
affect only the related CV (e.g. CV1), via the
main transfer function (e.g. G(1,1)), but also
many or all other CVs, via the interaction trans-
fer functions (e.g. G(i,1)).

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 11

If the impact of the interactions ( G(2,1) and 1.4.1 Intuitive Explanation of Predic-
G(1,2) in the example) is weak compared to the tive Control
main transfer functions (G(1,1) and G(2,2) in the
example), it might be possible to solve the mul- Controller detects future deviation
tivariable problem with two separate PID control- (e.g. violation of max. tolerable quality value)

lers (so called decentralized control). In some

cases, single interactions can be compensated
using lead-lag feedforward control (c.f. section SP

1.3.4). If however the impact of the interactions

is too strong (large gains, small time lags), or CV Controller already responds before deviation has occurred!

there are more than two or three interacting

variables, a real MIMO controller is necessary.
MV

The main principle of disturbance compensation

k+nc k+np
can be transferred from the SISO to the MIMO Past Future
case, and an MPC algorithm even facilitates the (measured values) (predicted values) Time t

implementation. A model of the influence of the

actual point k
disturbance variable to each CV is considered for in time
prediction of the CVs, such that the controller
can act in an anticipatory way against these Figure 10: Explanation of predictive control. Red: future
disturbance impacts. without control, green: future with control (planned
optimal trajectory), nc: control horizon, np: prediction
Although there are a lot of different MIMO con-
horizon
trol algorithms in theory, the model predictive
controllers dominate the field in industry.
Application Examples The controller observes and stores the process
• Quality control in distillation columns, see movements of the past. The controller can pre-
separate chapter 3. dict the future movements of the CVs in a cer-
tain time horizon (“prediction horizon”) using
• Temperature control of several adjacent
the internal complete dynamic process model
zones of furnaces with several burners, for
example, tunnel furnaces, glass melting including all interactions. The “future without
plants, glass feeders etc. control” assumes that all MVs are kept constant
at their actual value. The impacts of measurable
• Quality control in chemical reactors by ad- disturbance variables are already considered in
justment of reaction conditions such as the future without control.
pressure, temperature, feed/drain etc..
Moreover, the controller can ‘test’ (simulate)
• Vaporizers - for example drum steam gen- what will be the future impact of different
erators.
strategies to manipulate the process using the
• Mills - for example cement mills, sieve mills: available MVs (inside the control horizon). An
quality control (grain size) in combination optimization algorithm is applied to find the best
with flow rate maximization, manipulated MV strategy and the planned optimal trajectory
variables: sieve speed and mill feed. is displayed as ‘future with control’.
• Lime kiln (pulp mill): temperature control at The approach is similar to a chess computer:
hot end, cold end and hood draft; oxygen different combinations of future moves are ‘vir-
concentration control, via combustion gas tually played’ and evaluated according to their
feed, lime mud feed and fan speed. effects.
There are a lot of degrees of freedom in the
formulation of the optimization setup: future
control deviation, CV constraints and future MV
moves are included in the performance index
while MV limitations are hard constraints. Eco-
nomic targets can also be considered in the per-
formance index.
The optimization problem is solved for the whole
prediction horizon online at each sampling step
but only the first element of the calculated series
of MV moves is applied to the process. At the
next sample, the time horizon is shifted and the

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 12

complete optimization is re-started (‘moving (2) “Lean” MPC: If the constraints are not consid-
horizon principle’). ered during optimization, an analytical solution
of the optimization problem is possible, resulting
In other words: the control problem is formu-
in a straight-forward formula.
lated as an optimization problem and solved as
an optimization problem. The typical perform- This formula can be posted offline using per-
ance index looks like: formance index and process model. For the
r rT r r r r online calculation of the MVs, there are only
J = (w − y ) R(w − y ) + Δu T QΔu → min . some matrix multiplications to be performed (no
iterative search). This requires much less com-
w contains the time series of future setpoints, puting power than a full blown MPC. (MV con-
y contains the time series of future CVs (inside straints are fulfilled by the online-controller
the prediction horizon), anyway.) Such simplified predictive control algo-
rithms without online-optimization (lean MPC,
Δu contains the time series of future MV moves embedded MPC) and with a limited number of
(inside the control horizon). MVs and CVs can be implemented as a function
If the weightings in the Q diagonal matrix are block on DCS controllers.
increased, the controller moves its manipulated Example: The MPC function block from [1.] and
variables more cautiously resulting in a slower the ModPreCon function block from [2.] can
but more robust control action. Using the cope with up to four interacting MVs and CVs.
weighting factors in the R diagonal matrix, the Such an embedded MPC has the following ad-
relative significance of the individual controlled vantages:
variables can be specified. A higher weighting • The availability of the MPC function block is
(priority) for a controlled variable means that similar to a conventional PID controller.
this one moves more quickly towards the set- Backup strategies and arrangements for su-
point and remains more accurately at the set- pervision of communication to external PCs
point in steady state if it is not possible to are not necessary. All options of redundant
achieve all setpoints precisely. DCS hardware can be fully exploited, in-
creasing the availability of APC functions.
1.4.2 DCS embedded MPC versus
MPC with Online Optimization • The MPC function block can be connected to
other function blocks in the engineering
There are two completely different approaches system like any PID controller, using prede-
to solve the MPC optimization task, resulting in fined process tag types (templates for con-
two different classes of MPC algorithms: tinuous function charts).

(1) MPC with online optimization: If the optimi- • For operator control of the MPC function
zation problem is to be solved with respect to block there are standard faceplates.
constraints, a numeric iterative search for the • Look&Feel is similar to a conventional PID
optimum is necessary at each sampling step. controller and so familiar to plant operators,
Obviously, such an online optimization needs a reducing training effort. External consult-
lot of computing power. Therefore such types of ants as experts for special MPC software
MPC cannot be implemented in the controllers of packages are no longer needed.
a DCS, but need a separate PC interfaced to the
• Summing up, the starting prize for a turn-
DCS. Typically, an MPC is not manipulating the
key solution is reduced by an order of mag-
process directly but is commanding setpoints for nitude. This also means that small and me-
subordinate PID controllers, i.e. the MPC per- dium - sized applications become attractive
forms the coordination of several PID controllers for APC that does not allow amortizing a
in the basic automation level, in the sense of full-blown MPC.
“supervisory control”.
Typical examples of such full blown MPC-
software packages are e.g. DMC+ by AspenTech,
Inca by Ipcos [3.], ProfitController by Honeywell.

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 13

Figure 11: Embedded MPC on a PCS7 Operator Station

example), the corresponding partial transfer

function must be stabilized with a secondary
Cost
Full blown MPC
controller. For integrating processes, a simple P
ƒ Hardware controller (proportional component only) is
ƒ Software adequate as a secondary controller.
ƒ Engineering
Embedded MPC 1.4.3 Configuration of Predictive
ƒ Engineering Controllers
PID
At the core of any predictive controller there is a
ƒ Engineering Benefit
dynamic process model that describes the be-
haviour of the real process, i.e. the correlation
Figure 12: Cost benefit analysis of APC tools. For PID (incl. between all input and output variables, consider-
extensions) and embedded MPC, there are only engineer- ing all time lags. This model is typically identi-
ing costs, while full blown MPC additionally needs hard- fied from learning data, i.e. the main approach is
ware and software expenditure. similar to PID-Tuning (c.f. respective section
1.2.2):
• Excite the process with MV moves in manual
In some cases, an MPC function block is also mode of the controller, store learning data
helpful for a SISO process with particularly diffi- (using the CFC trend recorder for the PCS 7
cult dynamics. For processes with non-minimum MPC Configurator).
phase or strongly oscillating responses or with
• Identify process model: load data into MPC
very long deadtimes, it is better than PID control.
Configurator, select time span, filter noise if
Most MPC algorithms only work for stable proc- necessary. The identification itself is started
esses with a step response that settles to a fixed by a mouse-click, resulting in a fourth-order
value in a finite time. If the process is unstable plus dead time model for each part transfer
or includes an integrator (tank level control for

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 14

function. In case of 4 MVs and 4 CVs, this All control loops

PCS 7 APC-Tool
makes 16 part transfer functions. of plant
Manual action
• Verify model, i.e. simulate model with data CPM
User decision
different from the learning data.
OK
Performance?

Check
instrumentation

OK
Performance?

PID Tuning

OK
Performance?

Loops running well „APC candidates“

SISO MIMO
Dimension?
yes no yes no
Linearity? Linearity?

PID gain MPC model

MPC
scheduling scheduling
Figure 13: MPC Configurator in Simatic PCS 7 yes
Disturbance
measurable?
Feedforward dist.
compensation
The configurator automatically proposes a rea-
sonable controller sample time. During control- yes
Large deadtime?
ler design, the user can enter the following
specifications for the performance index: Smith
predictor
• Different weightings for different CVs, with
Loop pairing:
respect to their priorities in the specific ap- MV/CV assignment
plication. Example: process safety is more
important than product quality, product Figure 14: Typical procedure for selection of APC solu-
quality is more important than resource sav- tions. Blue shading refers to PCS 7 APC tools.
ings.
• Different MV move penalties for different
MVs. By increasing an MV move penalty, This description however has to be considered as
this MV will move more slowly, and the re- a rough framework: during individual applica-
lated process actuators will be treated with tions some steps can be skipped because prior
care. knowledge is available, or some intermediate
steps or iterations are required.
Before an MPC is applied to a real plant, it is
recommended to simulate the closed loop re- Some extensions to PID control like override
sponse using the process model identified. control, cascade control, ratio control or split-
range control are of structural nature: they are
not visible in Figure 14, but they are applied if the
respective structure is found in the setting of the
2 Typical Procedure to task.
If all trials in the SISO case fail, the allocation of
Improve Plant Per- MVs and CVs has to be checked, or actually there

formance using APC is a need for MIMO control (orange arrow). After
successful commissioning, APC functions as well

Tools as conventional control loops should be subject

to control performance monitoring (green ar-
For the improvement of plant performance using row).
APC tools, a typical procedure with several steps
is established.

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 15

2.1 Situation Analysis and product quality, for production costs (consump-
tion of raw materials and energy)? Which re-
Potential Estimation quirements for process automation are imposed
by chemical engineering (e.g. recipes for chemi-
2.1.1 Identification of Economic Tar- cal reactions), in order to ensure optimal produc-
gets for Process Control tion conditions? Which benefit could be achieved
by improved control performance?
The starting points for any discussion are eco-
nomic considerations and the requirements for Table 1 summarizes the typical objectives for
process control imposed by chemical engineer- APC solutions in (predominantly continuously
ing: Which measurable variables are relevant for operated) process plants (following [4.], p. 201).
the economic success of process operation?
Which of them are relevant for plant safety, for

Table 1: Objectives for APC solutions

Objectives Aspects
Productivity and Increase throughput
Profitability
Minimize energy consumption
Reduce changeover times, e.g. for changes of operation mode, raw material or
target products (grade changes)
Increase yield
Reduce processing time
Quality Increase repeatability; make process operation more steady and smooth
Minimize variance of quality parameters
Reduce effort for chemical analyses
Reduce production of giveaway or inferior qualities
Operability and Increase tolerance against fluctuations in raw material
Availability
Reduce sensitivity to disturbances
Increase plant uptime
Avoid plant shutdowns and reduce shutdown times
Increase plant availability, flexibility and robustness
Operator Control Cope with change of operators/shifts
Reduce operator workload
Increase operating comfort
Safety Increase occupational health and safety
Increase process and operational safety
Ecology Minimize environmental impact, pollution and wastage
Minimize emissions
Save effluents and sewage

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 16

tuned as well as possible, for example by appli-

Constraint
cation of a PID-Tuning-Tool (cf. section 1.2.2).
Constraint
This is the way to avoid ‘breaking a butterfly on a
wheel’ – if a control problem can be solved by
simple means, there is no need to go for elabo-
rate APC methods.
New SP
In the framework of an APC concept, supervisory
PV controllers, like multivariable MPC, are typically
established as master controllers for many PID
slave controllers. This is a cascade structure and
in cascade structures it is important to avoid
Old SP
modifications of the tuning of slave controllers
after the tuning of the master controller is fin-
Time ished because the slave closed loops are part of
the (time-invariant) process model used in the
master controller. This is one more reason to
Figure 15: “Constraint pushing” – make the most of the
economic potential of a plant by improved process control finish the optimization of the basic PID control
with reduced CV variance: move setpoints closer to CV loops before the implementation of supervisory
constraints. APC functions and is in line with the general
procedure for commissioning of cascade control
structures ‘starting from inside loop, proceeding
In many cases, such economic objectives are with outside loop’ i.e. first the slave controller,
reached in two steps: then the master controller.
1. Reduction of CV variances by improved
process control. 2.2 Concept Definition of
2. The reduced variances allow moving the APC-Functions
setpoints closer to critical constraints with-
out the risk of frequently violating the con- The most important questions to be answered in
straints. This is called “constraint pushing”: the context of an APC concept are:
make the most of the process by using the • Which are the objectives of the APC solu-
full physical capabilities of the plant e.g. to tion?
maximize throughput or minimize energy • Which MVs and CVs are available for the APC
consumption. solution?
2.1.2 Analysis of Current Plant Status • Which algorithms are to be used?
incl. Control Performance Mo-
• How to design the structure of the control
nitoring system?
A thorough analysis of the current status of the 2.2.1 Selection of Appropriate Ap-
plant is required before starting to think about
APC. The status of instrumentation (sensors,
proach Based on Analysis
actuators) and automation (especially the exist- The basic questions with respect to selection of
ing basic control loops e.g. PID controllers) the appropriate approach are: SISO or MIMO
should be checked. If a control performance control? Linear or nonlinear control?
monitoring system (cf. section 1.2.1) is estab-
lished, typical problem areas can be easily found, 2.2.1.1 SISO or MIMO Control?
e.g. loops that cause a lot of alarms; loops that
are frequently running in manual; loops that With the background of the respective section
show extremely high CV variances or are perma- 1.4 a checklist for the decision SISO or MIMO
nently oscillating. control could be the following:
• Does the plant unit (e.g. apparatus, ma-
2.1.3 Review of Basic Automation chine, plant section) contain several CVs?
then PID-Tuning • Are there significant interactions between
Errors in Instrumentation found during the different control loops? Test: If a setpoint
analysis must first be corrected, e.g. bad calibra- step is commanded for one loop, are there
side-effects in other control loops? Any
tion of sensors; excessive friction of valves; leak-
problems in tuning single loops, because
ages in compressed air pipes etc. Afterwards it
modified control parameter sets in one loop
must be ensured that all PID controllers are have impacts on other control loops?

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 17

• Are the variables having an impact on

neighbouring loops really moving during
plant operation? Counter-example: the cor- If one of the issues mentioned above require a
relation between pressure and temperature nonlinear control concept, there are the follow-
of a gas volume in a reactor or vessel can be ing solutions possible that can be examined in
neglected if either pressure or temperature the following order:
is kept constant during plant operation.
2.2.1.2.1 Compensation Functions in-
• Is the control performance of the CVs suffer- between Controller and Process
ing from interactions really relevant for the
economic success of plant operation? Example: The effect of a nonlinear valve signa-
Counter-example: although there is a physi- ture can be compensated by a polygon function
cal correlation between the feed flow into a block between the MV output of the controller
vessel and the fluid level inside the vessel, in block and the input of the valve block. In this
many cases it is sufficient for level control to case, the transformation of the MV limits must
manipulate the outflow valve because very be considered.
precise level control is not required from a
chemical engineering point of view. In a similar way, the effect of nonlinearity at the
process output (e.g. sensor characteristic) can be
• Any attempts failed in the past to solve in-
compensated for by a polygon function block in
teraction problems using lead-lag feedfor-
front of the CV input of the controller. In this
ward control (cf. section 1.3.4)? Spent ex-
case the related setpoint must be transformed in
cessive effort for design and engineering of
decoupling structures in similar plant units? the same way.

If in doubt, a MIMO process model can be identi- In both cases, the compensation functions be-
fied by the MPC Configurator. In the matrix of come part of the controlled process from the
transfer functions, the step response of the in- perspective of the controller. The aim is always
teractions can be observed and compared to the to make the overall response of the controlled
main transfer functions on the matrix diagonal. system (consisting of process and compensation
elements) as linear as possible.
Tips for the decision to apply smith-predictor or
lead-lag feedforward and the discrimination with 2.2.1.2.2 Operating-Point-Specific Parame-
respect to a SISO MPC can be found in the re- ter Sets
spective sections 1.3.3 and 1.3.4.
For SISO case: PID gain scheduling according to
section 1.3.2.
2.2.1.2 Linear or Nonlinear Control?
For MIMO case: MPC multi-model control, cf.
Most process plants are nonlinear by nature
example project in PCS 7 Advanced Process Li-
(physics, chemistry), i.e. the impact of a certain
brary. This approach is related to the basic idea
MV move is different depending on the actual
of operating-point-based parameter control with
operating point of the process. Typical causes
PID controllers. Since the model parameters of
are nonlinear valve characteristics; nonlinear
the ModPreCon block cannot be modified at
reaction kinetics; phase transitions
runtime, the schedule for selecting the suitable
(solid/fluid/gaseous); titration characteristics etc.
parameter sets becomes a schedule for selecting
To decide on a linear or nonlinear control con-
the suitable models.
cept, the following questions are relevant:
• Is an existing linear control concept (e.g. Several ModPreCon instances with different
conventional PID control) working well at a models for different operating points run at the
certain operating point, but oscillating or same time in parallel. The local optimal models
very sluggish at different operating points are determined by exciting the process at the
(e.g. after load change)? various operating points with small amplitudes,
• Does the operation of a process plant re- so that only the reaction of the nonlinear proc-
quire frequent changeovers of operating ess in the ambience of this operating point is
point, e.g. in the context of batch recipes; in registered.
the context of grade changes (multi-product
The final manipulated value for each manipu-
plant); in the context of a flexible produc-
lated variable is formed as a weighted mean
tion (load changes)? Counter-example: a lin-
ear controller can cope even with a strongly value of the manipulated variables proposed by
nonlinear process if a continuous plant is the different controller instances. The weighting
running always at the same constant operat- factors 0 ... 1 are formed in the same way as the
ing point. membership functions known from fuzzy logic
so that the sum of all weights is always one and

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 18

each controller has the highest weighting at its multivariable control problem increases
own operating point. quadratically with the number of MVs and
CVs (‘Divide and conquer!’).
2.2.1.2.3 Trajectory Control • The quality of measurements is also relevant
This approach proposed by [11.] neatly com- for the selection of CVs: are the measure-
bines the advantages of an open loop controller ments reproducible, available in real time
(feedforward control) with those of a closed and low in noise?
loop controller (using process value feedback). • Which MVs are selected because they have a
The controller follows a previously optimized significant and direct impact on the desig-
trajectory of setpoints and manipulated vari- nated CVs? Normally, it is recommended to
ables; in other words, it only needs to compen- search for as many MVs as there CVs to cope
sate small deviations between the stored trajec- with, resulting in a ‘quadratic system’. The
tory and the current plant state. A trajectory is controller can then bring all CVs exactly to
an optimum series of manipulated variables over their target values as far as allowed by con-
time and the process values that match them. straints. Otherwise you have constituted a…
The required manipulated variables are read into • non-quadratic control problem (number of
the ModPreCon block via the inputs MV1Traj to MVs is not equal to number of CVs). If there
MV4Traj and added to the values of the manipu- are less MVs than CVs, or some of the MVs
lated variable calculated by the algorithm (in have reached their limits, there is a lack of
automatic mode only). Among other things, the degrees of freedom, and it is no longer pos-
advantage of this is that the effective manipu- sible to reach all setpoints exactly. The MPC
lated variable acting on the process (the sum of algorithm then finds a compromise that can
trajectory and controller action) is limited as be influenced by the selection of CV weights
configured in the controller block. The process (priorities) in the MPC Configurator and CV
values from the trajectory are linked to the cor- zones in the MPC function block: CVs with
higher priority will have lower control devia-
responding setpoint inputs SP1 to SP4 of the
tions with respect to their setpoints or
controller. As long as the process reacts exactly
zones. If there are more MVs than CVs or if
as planned in the trajectory, it will respond to
some of the CVs are already within their
the series of MV values from the trajectory with control zones, there is surplus freedom in
the corresponding series of CV values and the the control problem. A lean predictive con-
control deviation is zero. It is generally known troller algorithm (cf. respective section
that a non-linear dynamic process can linearized 1.4.2) however, cannot recognize this situa-
around a fixed operating point or a steady state tion explicitly and use the free manipulated
of the system but it is also possible to linearize it variables for optimization. The MPC block
around a trajectory. therefore moves all manipulated variables to
values that meet the targets in terms of con-
2.2.2 Design of Control System trolled variables and then leaves them there.
Structure, Selection of MVs, In the context of unipolar actuators (e.g.
CVs and Constraints heating and cooling) it makes sense to com-
bine two unipolar actuators into one bipolar
For SISO controllers, a one-to-one assignment of actuator using a split-range function block.
MVs and CVs must be specified: for each control-
ler it must be decided which MV is used to ma- 2.2.2.1 Hard- and Soft Constraints
nipulate a certain CV. If several MVs are in ques-
tion, select the MV that has the strongest impact MV limitations are ‘hard constraints’ that always
on the CV (in the sense of gain) and the most have to be accepted by the controller. On the
direct impact (in the sense of small time lag) other hand, CV constraints (control zones) are
‘soft constraints’: violations of such constraints
For an MPC concept, such an assignment is not are regarded as control deviations and avoided
necessary because the MPC makes use of all MVs as much as possible. However, there can actually
to manipulate all CVs. For selection of MVs and be a temporary violation of a CV constraint,
CVs in a multivariable control system, the follow- similar to an overshoot with respect to an exact
ing issues are to be considered: setpoint.
• Which CVs are interesting in the sense of the
arguments of section 2.2.1.1? Which CVs are In multivariable controllers, it is advisable to
relevant for plant safety, product qualityand make use of the fact that from the perspective of
economics? Which requirements for control the application, only some of the controlled
precision are imposed by chemical engineer- variables need to move to a specified setpoint
ing? Avoid inflating the problem more than exactly while others only need to remain within
necessary because the complexity of the a defined range. A typical example would be

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 19

quality characteristics for which a tolerance band During the recording of learning data, significant
is specified. In principle, the effect of the control disturbance events must be avoided such as load
zones of an MPC is the same as the effect of a changes; grade changes of raw materials; set-
dead band in a PID controller but extends over point steps of neighbouring control loops; main-
the entire future prediction horizon. In other tenance activities in the plant; significant modi-
words, if, for example, the predicted controlled fications in upstream plant sections or in the
variable CV1 in the entire prediction horizon is mains power or steam supply, etc.
within the band SP1 ± SP1DeadBand, the con-
There are different possibilities for the choice of
troller sees no reason whatsoever to change any
the excitation signal. For PID tuning a single step
manipulated variable. While a dead band in a PID
response is normally sufficient. For the identifi-
controller tends to put stability of the control
cation of MIMO processes, all MVs must be ex-
loop at risk, CV zones in individual control chan-
cited, ideally one after the other. Signals sym-
nels of an MPC generally relieve the multivari-
metrical to the operating point are preferred, i.e.
able controller overall. Using CV zones, the ac-
small steps upwards and downwards from the
tion of a soft override control can be achieved. If
operating point. The amplitude of the excitation
one CV must be limited only in one direction,
must be carefully specified and discussed with
e.g. prevented from violating an upper limit, the
facility and plant operators:
exact setpoint is not relevant. Setpoint and zone
• The amplitude should be large enough to
width are then specified such that the lower
make sure that the process response is
limit is never reached during operation. clearly visible relative to measurement noise
(signal-to-noise ratio).
2.3 Configuration of APC-
• It must not be too large to avoid process
Functions nonlinearities having a strong impact.
After selection of an appropriate control algo- • It should be not bigger than necessary to
rithm and design of the control system structure, avoid distraction of running production.
the next step is the configuration, i.e. the pa-
• In the MIMO case, the process should be
rameterization and commissioning of the control excited uniformly with respect to all CV
function. For PID controllers and MPC functions channels, i.e. MVs with weak impact must
there are dedicated software tools like e.g. PID- be excited with bigger amplitudes.
Tuner and MPC-Configurator.
Note: the routine inspections related to control-
The principal procedure however is the same for ler commissioning (cf. section 2.3.4) have to be
all controller types that rely on a model of proc- performed before recording of learning data in
ess dynamics for controller design. Such a the context of computer aided control system
model must be identified from learning data design.
because it is not usually known beforehand.
2.3.2 Modelling
2.3.1 Process Excitation and Re-
cording of Learning Data The identification of the dynamic process model
from learning data is performed in most of the
In order to identify the process dynamics from respective software tools automatically without
learning data, the process must be stimulated to prompting the user for dedicated specifications
move i.e. actively excited via the MVs. If at least of a model structure. Simple functions for data-
a stable and not too oscillatory (although sub- pre-processing are sometimes offered e.g. selec-
optimal) controller already exists, the excitation tion of a relevant time slice from a longer ar-
can be achieved by setpoint steps in automatic chive data set, or low-pass filtering for reduction
mode; otherwise MV moves in manual mode are of measurement noise.
necessary.
Verification of the model is generally very impor-
It is important to perform the experiment at tant, i.e. comparison of a simulation of the iden-
exactly the same operating point and under tified model to real measured data, in order to
exactly the same conditions that are designated check if the model response sufficiently fits the
for operation of the controller in order to really response of the real process. Ideally, different
observe the process behaviour relevant for con- excitation signals are used for verification and
trol. Before starting the experiment, a steady learning data. Moreover, it is recommended to
state of the process is required, because tran- check the plausibility of the model at a descrip-
sients that have not yet faded away and have tive level, e.g. using model parameters or model
been caused by phenomena outside the learning step responses: are gains and time constants in
data will spoil the estimation results.

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 20

the order of magnitude expected from physical • if the sign of the controller gain (of PID
conceivabilities or other a priori knowledge? controller) fits to the sign of the process
gain
2.3.3 Controller Tuning
• if additional calculations for pre- or post-
Similar to model identification, controller design processing of controller signals (e.g. unit
is performed automatically by the related soft- transformations, measurement corrections,
ware tools. The specifications for controller de- linearization functions according to section
sign naturally differ for different controller 2.2.1.2.1 ) are working properly.
types.
Frequently (hopefully), these checks have been
performed before recording learning data (c.f.
2.3.3.1 PID-Tuner section 2.3.1).
In PID-Tuner, the user can specify the controller Typically, during commissioning the process is
structure: P-, PI- or PID-controller. PI control is driven to the operating point in manual mode of
applied in most cases. P controllers are suitable the controller and at the operating point, a
especially for integrating processes like e.g. level bumpless switchover from manual to automatic
control. The differential action of a PID controller mode is performed. If the controller does not
allows faster control response and better com- respond as expected, it is immediately switched
pensation of disturbances if stronger movements back to manual mode. In very critical processes,
of the actuators are accepted (wear, energy MV limitations are limited to a very small zone
consumption). around the current MV value before the control-
PID controllers principally allow modifying the ler is switched to automatic mode for the first
proposed parameters manually (fine-tuning) and time.
many tuning tools allow study of the response of Only MPC allows running the controller ‘pas-
the closed control loop in a simulation before sively’ in automatic mode without connecting its
new parameter sets are downloaded into the MVs to the process. This way, the MV proposals
target system. of the MPC can be checked for plausibility and
eventually typed in manually: ‘advisory control’,
2.3.3.2 MPC-Configurator ‘human in the loop’.
The response of a predictive controller is mainly In contrast to that, a PID controller is not allowed
determined by the process model used inside the to run in automatic mode without connection of
controller. Internal parameters like prediction its MV to the process because this would cause
horizon and control horizon can be specified integrator windup.
automatically based on the process model.
If the control loop is running stably in automatic
The user only has to specify CV priorities and MV mode, the concrete requirements with respect to
move penalties in the performance index as control performance are checked in detail. A first
described in section 1.4.3. Manual tuning of impression of control performance gain can be
other controller parameters is not possible for obtained by a small setpoint step. In the MIMO
MPC, just the simulation of the closed control case too, setpoint steps are initially tested at
loop. single CVs, and the (undesired) ‘crosstalk’ to
neighbouring control channels is observed as
2.3.4 Commissioning
well. If possible, certain operating procedures or
If the closed loop response has been already disturbance scenarios relevant for control opera-
checked in simulations, the risks for controller tion are played through and tested e.g. load
commissioning are reduced. Before really closing changes; grade changes of raw materials; limita-
the loop, you should generally check… tions in heating or cooling system etc. Other-
• if the measured CVs arrive correctly at the wise, the disturbance response can be tested by
controller in the physical unit of the setpoint an artificial MV step. Most controller function
specification blocks offer an input variable for offsets to the
• if the MVs have an impact on the process at MV.
all, i.e. if the CVs respond to small MV steps A control performance monitoring system ac-
in manual mode cording to section 1.2.1 can be initialized after
• if the specification of MV limits fits to the successful controller commissioning.
process actuator

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 21

2.4 The Most Important Process plants are subject to permanent modifi-
cations, like a living organism. Modifications
Rules at a Glance result both from undesired aging processes like
catalyser deactivation or heat exchanger fouling
• First check instrumentation and then tune and from conscious measures of maintenance,
controllers. modification or expansion of the plant in the
• For cascade structures, first tune the slave operational phase. This implies that the static
controller and then the master controller. and dynamic behaviour of a process plant is
subject to changes - slower or faster, smaller or
• First optimize the basic control loops and bigger. You must be aware of the fact that the
then design APC functions. process models used for APC in the course of
• If several MVs come into consideration, time will diverge more and more from reality,
select the MV with the strongest impact (in implying negative effects on control perform-
the sense of gain) and the most direct im- ance. Maintenance of an APC solution requires
pact (in the sense of small time lag). permanent observation and evaluation and in
this context the tools for control performance
• The general aim is to design the overall monitoring described in section 1.2.1 are help-
response of controlled plant and compensa- ful. Suggested actions are scheduling of an ap-
tion blocks as linearly as possible. propriate date for re-identification of some or all
process models; eventual re-configuration of the
APC solution (include new MVs or CVs, or elimi-
2.5 Documentation, Train- nate MVs or CVs that are no longer necessary or
available); re-commissioning and last but not
ing and Maintenance least, the initialization of follow-up projects.
It is good practice in the sense of benefit control
to compare the performance of an APC solution
to the performance of the automation system
3 Case Study Distilla-
existing beforehand (conventional control or
manual operation) and to document the differ-
tion Column
ence. Distillation is one of the most important separa-
tion processes in the chemical-pharmaceutical
In order to ensure the long term benefit of ad-
industry. It is a thermal separation process to
vanced control, documentation, training and
separate a mixture of several fluids dissolvable in
maintenance play an important role. A func-
each other, making use of the different relative
tional documentation of the APC solution should
evaporation and boiling points. Typical applica-
allow usage and maintenance of the solution,
tion areas of distillation are alcohols or crude oil.
even for people who have not participated in the
There are two main classes: (1.) batch distilla-
design of the solution. Especially in the context
tion, and (2.) continuous distillation (e.g. rectifi-
of a first-time application of APC in a facility,
cation, extractive or reactive distillation.). In
confidence-building measures are very impor-
rectification (counter flow distillation), a mixture
tant to motivate plant operators, process and
of several components is separated in at least
automation engineers and facility managers.
two fluid streams. At the column head, the head
Essentially, the attitude ‘rather a bad solution
product (low-boiling component) is extracted,
than a solution that I don’t understand’ must be
and at the column bottom, the bottom product
overcome. The functionality of most APC solu-
(high-boiling component).
tions can be described in an illustrative way like
in chapter 1 of this document, and as can be Due to the strong thermodynamic interactions
seen from section 1.4.1, even for MPC, a basic inside the distillation tower, conventional PID
understanding of the approach can be achieved control technology can only control the product
without diving deeply into the mathematical quality (concentration, or substituted by tem-
theories. Formal and on-the-job training have to perature) either at the head or at the bottom of
be considered separately. In spite of the required the column. Therefore, up to now there are 5
effort, it is recommended to offer different train- different alternative structures for distillation
ing for different target audiences with different column control via one single temperature [12.].
focus and different coverage.
Thanks to the routine application of a MIMO
To this end, it is helpful to provide a simulation controller with two MVs and two CVs, now for
environment not only in the design and training the first time all columns that have been driven
phase of an APC project but also in the operating with one of the five conventional control struc-
phase. tures up to now, can be equipped with the same

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 22

standard automation structure: [10.]. This stan- application can be specified using the CV priori-
dard structure comprises a complete quality ties in the performance index of the MPC.
control: control of head- and (!) bottom tem-
perature via reflux and heating vapour flow. The
relative importance of the CVs for an individual

Figure 16: P&I diagram for a rectification column with quality control via MPC. (Source: OS picture of the PCS 7 solution tem-
plate described in [10.] . The measured values are greyed, because here they are in status simulation.)

3.1 Template-based Imple- controllers), the usual guidelines for the engi-
neering of cascade control structures have to be
mentation considered as shown in the CFC template cas-
cade control:
The CFC template (process tag type) for the pre- • The actuation range of the primary control-
dictive controller is the starting point for the ler (in the respective MV channel) must
implementation of the multivariable controller. match the setpoint range of the secondary
This signal flow chart is modified at the output controller to ensure proper operation of the
side: the output variables of the predictive con- anti-windup functions of the primary con-
troller are not directly connected to the plant but troller.
are connected as external setpoints to the exist-
• If the secondary controller is not operated in
ing PID flow controllers for reflux and heating
cascade mode (automatic mode with exter-
steam. nal setpoint) but in any other mode (for ex-
Because this is actually a cascade structure (with ample manual or automatic mode with local
one MIMO master controller and two PID slave setpoint) and therefore does not respond to

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 23

commands of the primary controller, the ventions and operator workload are reduced
corresponding channel of the primary con- accordingly.
troller must be put into tracking mode. The
• Feed flow as the main measurable distur-
manipulated value of the primary controller
bance variable is used to improve prediction
tracks the process value (or setpoint) of the
accuracy, in order to compensate for feed
corresponding secondary controller to allow
a bumpless return to cascade mode. The dif- flow fluctuations in a predictive way by ad-
ference between tracking the setpoint and justing reflux and heating steam without
tracking the process value becomes appar- tolerating significant product quality devia-
ent when the secondary controller is put tions from setpoint. Control performance in
into manual mode. If the process value is this respect can be improved significantly
tracked, the response is similar to the ‘Track compared to conventional PID control.
setpoint to process value in manual mode’ • The main economic benefit however can be
of a simple controller. achieved according to Figure 15 by a dedi-
• Commissioning of cascade control structures cated move of the quality control setpoints.
The constraints for allowed head and bot-
is performed from inner loop to outer loop,
tom temperatures result from specifications
i.e. first all slave controllers are optimized
of product quality, considering actual vari-
with the PID tuner and afterwards the mas-
ances of controlled variables. For optimal
ter controller is tuned by the MPC configura-
operation of the column with minimal en-
tor.
ergy consumption (i.e. steam consumption),
Due to the nonlinear thermodynamic effects in the head temperature setpoint is selected as
the distillation column, quality control with a high as possible and the bottom setpoint as
linear MPC is only feasible in a limited range low as possible, i.e. the temperature span
around a defined operating point. Therefore it is inside the column is minimized. In other
important to do experiments for recording learn- words, do not produce the best quality
ing data at this operating point and to restrict achievable, but exactly this acceptable qual-
ity that can be produced with minimal costs.
the MV range of the MPC for automatic mode to
In other plant constellations there can be
the validity range of this linear model. The MV
other optimization targets. For example,
limits for manual mode however should span the
maximize throughput by increasing feed
whole SP range of the slave controllers in order flow step by step under quality control until
to allow start up and shutdown of the process in heating steam flow is settling close to the
manual mode of quality control. upper limit.
When using the standard CFC templates of the
PCS7 APC library, control performance monitor-
ing according to section 1.2.1 for all PID loops
and the MPC is automatically included in the
project.

3.2 Economic Benefit

What is the economic benefit of the APC solution
‘quality control by MPC’ in the context of distilla-
tion columns?
• Now for the first time it is possible to specify
the quality of head and bottom product in-
dividually and control it in closed loop. The
variance of the quality parameter, which
was not controlled at all before, can be dra-
matically reduced; the variance of the qual-
ity parameter, which was PID controlled be-
fore, can still be significantly reduced. Over-
all process control is more even and smooth
with product quality more exactly adjust-
able. Smaller variances of product quality
typically allow for savings in downstream
processing units, e.g. reduced consumption
of additives.
• Automatic quality control needs less opera-
tor attention, i.e. number of manual inter-

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 24

4 Literature
[1.] Siemens AG, Automation and Drives: Online-
Hilfe zur PCS7 APC-Library V7.0 SP 1, Nov. 2007.
[2.] Siemens AG, Sektor Industry: Online-Hilfe
zur PCS7 Advanced Process Library V7.1, Nov.
2008.
[3.] Siemens AG, Sektor Industry: PCS7 AddOn-
Katalog. (available at
www.automation.siemens.com )
[4.] Dittmar, R., Pfeiffer, B-M.: Modellbasierte
prädiktive Regelung - Eine Einführung für Inge-
nieure. Oldenbourg Verlag, München, 2004.
[5.] Dittmar, R., Pfeiffer, B-M.: Modellbasierte
prädiktive Regelung in der industriellen Praxis.
at 12/2006, S. 590-601.
[6.] Pfeiffer, B-M.: Control Performance Monito-
ring mit einer Kombination aus stochastischen
und deterministischen Merkmalen. GMA-
Kongress 2005, Baden-Baden. VDI-Berichte
1883, VDI-Verlag, Düsseldorf, S. 411-420.
[7.] Föllinger, O.: Regelungstechnik. 6. Auflage,
Figure 17: Distillation columns
Hüthig-Verlag, Heidelberg 1990.

The overall benefit of an APC solution for a distil- [8.] Preuß, H.-P., Linzenkirchner, E., Kirchberg,
lation column (like that described e.g. in [13.]) K.-H.: SIEPID – ein Inbetriebsetzungsgerät zur
typically adds up to more than €100,000 per automatischen Regleroptimierung. Automatisie-
year. rungstechnische Praxis atp 29(1987), Heft 9, S.
427-436.
Typical empirical values for the benefit of APC
solutions are: [9.] Pfeiffer, B-M.: Standardisierung gehobener
Regelungsfunktionen als Messstellen-Typen.
• Throughput increased by 1…5%
GMA-Kongress 2007, Baden-Baden. VDI-Berichte
• Yield increased by 2..10% 1980, VDI-Verlag, Düsseldorf, S. 83-94.

• Energy consumption reduced by 3…10% [10.] Pfeiffer, B-M., Lorenz, O.: Unit-orientierte
Musterlösungen für Advanced Control (Unit-
• Standard deviation of process variables
oriented solution templates for advanced
reduced to 25…50% of the value before.
control) - Beispiel Destillationskolonne (Example
distillation column). Automation 2008, Baden-
Baden. VDI-Berichte 2032, S. 11-14, VDI-Verlag,
3.3 Summary Düsseldorf.

APC methods are a tool of vital importance to [11.] Pfeiffer, B-M.: PID-Regelung von Batch-
improve plant performance with respect to pro- Prozessen entlang vorab optimierter Trajektorien
ductivity and economics, product quality, oper- – Studien am Chylla-Haase Reaktor-Benchmark.
ability and availability, agility, safety and envi- atp 4/2003, S. 76-86.
ronmental issues. [12.] Kister, H.Z.: Distillation Operation. Mcgraw-
APC solutions can be realized much more cost Hill Professional, 2007.
effectively due to a DCS embedded implementa- [13.] Schüler, M.: Implementing MPC at Ineos
tion with standard function blocks and pre- Phenol’s Gladbeck site. Hydrocarbon Engineer-
defined CFC templates as offered by Siemens in ing, Vol. 11, No. 10, Oct. 2006.
the APC library. [14.] VanDoren, V.: Advances in control loop
optimization. Software takes users from simple
tuning to plant-wide optimization. Control Engi-
neering May 2008.

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d
White Paper | SIMATIC PCS 7 APC-Portfolio | October 2008 25

www.siemens.com Siemens AG

All rights reserved. All trademarks used Sektor Industry, IA AS SM MP7

are owned by Siemens or their respective owners. Östliche Rheinbrückenstr. 50
D-76181 Karlsruhe
© Siemens AG 2008

A white paper issued by: Siemens, Sector Industry, IA AS SM. © Siemens AG

2008 All i h d