Beyond Simulation: Computer Aided Control System Design Using

Equation-Based Object Oriented Modelling for the Next Decade

Francesco Casella1 Filippo Donida1 Marco Lovera1

1
Politecnico di Milano, Dipartimento di Elettronica e Informazione, Italy
{casella,donida,lovera}@elet.polimi.it

Abstract lot of activity in this field, which eventually culminated in

After 20 years since their birth, equation-oriented and the development of the Modelica Language [32].
object-oriented modelling techniques and tools are now At the beginning of that decade, papers appeared on the
mature, as far as solving simulation problems is concerned. subject in the IEEE’s Control Systems Magazine [31, 10],
Conversely, there is still much to be done in order to pro- which discussed the potential of OOM for control sys-
vide more direct support for the design of advanced, model- tem design. Reading those papers in retrospect shows that
based control systems, starting from object-oriented plant some of the promises where actually met or even exceeded:
models. Following a brief review of the current state of OOM is now a mature field, both from a theoretical side
the art in this field, the paper presents some proposals for and from the point of view of available simulation tools.
future developments: open model exchange formats, auto- On the other hand, much work still has to be done on
matic model-order reduction techniques, automatic deriva- two fronts. The first one, which has a more “political”
tion of simplified transfer functions, automatic derivation nature, is spreading the OOM culture among in the con-
of LFT models, automatic generation of inverse models for trol engineering community, which is still largely domi-
robotic systems, and support for nonlinear model predictive nated by block-oriented modelling, and by the (mis)use of
control. Matlab/Simulink for physical systems modelling; this chal-
lenge is of paramount importance, but it out of the scope
Keywords Control system design, symbolic manipula- of this paper. The second one, instead, is to develop tools
tion, model order reduction, CACSD. which allow to use EOOL models and tools not only for
simulation, but also for the design of advanced control sys-
1. Introduction tems. The availability of such tools is crucial in order to
narrow the gap between the large body of highly sophis-
Control system engineering requires to master the dynam- ticated control theory developed during the last 20 years,
ics of plants which are in general complex, interacting, and the application of this theory to real-life cases, beyond
multi-physics and multi-disciplinary. This explains why textbook-sized examples. This is the topic of the present
object-oriented modelling (OOM) and a-causal, equation- paper.
based, object-oriented languages (EOOL) always had a Given the background and the past experience of the
very strong connection with control system design. It is authors, the discussion might be biased towards the Mod-
by no means accidental that much pioneering work in the elica language and related tools. However, strictly object-
OOM field was carried out within systems and control oriented features such as inheritance, encapsulation and hi-
departments and research groups: consider, for example, erarchical composition do not play any significant role in
the Omola language and the associated OmSim simula- the analysis and proposals made within this paper, which
tion environment, developed at the Department of Auto- essentially focuses on transformations of flattened models.
matic Control of Lund Technical University [29, 30, 4], or On the contrary, the discussion is relevant for any equation-
the MOSES environment developed at the Dipartimento di based modelling language, provided that it is a-causal and it
Elettronica of Politecnico di Milano [26, 9]. During the ’90, allows symbolic manipulation of the equations by the com-
OOM was considered a very promising tool for Computer piler.
Aided Control System Design (CACSD), and there was a The paper is structured as follows: Section 2 gives a
high-level view of the modelling activities required for con-
trol system design, while the following Section 3 discusses
how currently available tools can help the control engineer
2nd International Workshop on Equation-Based Object-Oriented
Languages and Tools. July 8, 2008, Paphos, Cyprus. in his/her task, with particular reference to Modelica tools.
Copyright is held by the author/owner(s). The proceedings are published by Sections 4 and 5, which are the core of the paper, pro-
Linköping University Electronic Press. Proceedings available at:
http://www.ep.liu.se/ecp/029/
pose several research and development directions to sub-
EOOLT 2008 website: stantially increase the level of support to the control en-
http://www.eoolt.org/2008/

35
gineer, willing to apply advanced control theory to real-life In many cases, the dynamics of systems in the form (1) is
problems. Section 6 concludes the paper with final remarks. approximated by (2) via linearization around some equilib-
rium point. There is also a vast body of advanced control
2. The role of mathematical models in techniques which are based on discrete-time models:
control system design
x(k + 1) = f (x(k), u(k), p, k)
The design of control systems always requires some knowl- (4)
y(k) = g(x(k), u(k), p, k)
edge about the dynamic behaviour of the plant under con-
trol. When the plant design is mature and well-known, where the integer time step k usually corresponds to the
and the control system design is based on Proportional- sampling time Ts of a digital control system. Many tech-
Integral-Derivative (PID) controllers, the latter is often niques are available to transform (1) into (4).
based on past experience and possibly on some empirical These models must capture the fundamental dynamics
measurements. In this case, which covers the vast majority which is relevant for control system performance, while
of installed industrial controllers, no (explicit) dynamical remaining as simple as possible: most advanced control
modelling is needed. design techniques start to become intractable for systems
On the other hand, in an increasing number of cases, the of order greater than about ten. If the models are simple
performance of the control system is becoming a key com- enough, it is also sometimes possible to express the de-
petitive factor for the success of innovative, high-tech sys- pendence of key dynamic features (such as, e.g., the natu-
tems. To name a few examples, consider high-performance ral frequency and damping coefficient of an oscillating dy-
mechatronic systems (such as robots), vehicles enhanced namics) from plant design data. This can be very important
by active integrated stability, suspension, and braking con- to assess the impact of physical system design decisions
trol, aerospace system, advanced energy conversion sys- on controller performance. For example, if the natural fre-
tems. All these cases possess at least one of the following quency of the first mode of oscillation limits the controller
features, which call for some kind of mathematical mod- bandwidth, and it is found that this frequency mainly de-
elling for the design of the control system: pends on the stiffness of a certain mechanical component,
then it might be reasonable to change the mechanical de-
• closed-loop performance critically depends on the dy-
sign of that component in order to improve the overall per-
namic behaviour, which is not well-known in advance;
formance.
• the system is complex, made of many closely interact- In order to derive such simple models, it is usually
ing subsystems, so that the behaviour of the whole sys- necessary to introduce many, sometimes drastic, simplify-
tem is more than just the sum of its parts; ing assumptions: all those phenomena that only marginally
• advanced control systems are required to obtain com- affect the equilibrium values and/or the control-relevant
petitive performance, and these in turn depend on ex- dynamics of the system are neglected. This activity re-
plicit mathematical models for their design; quires highly skilled and experienced modellers, with a
good knowledge of control design techniques, as well as
• the system is very expensive and/or safety critical, re-
of domain-specific strategies for model simplification.
quiring extensive validation of good control perfor-
mance by simulation. 2.2 Detailed models for system simulation
In most of these cases, two different classes of mathemati- At the other end of the modelling spectrum, detailed sim-
cal models are derived: compact models for control design ulation models can be found. Although it is always neces-
and detailed models for system simulation. sary to make reasonable modelling assumptions (a model
is always a focused and limited description of the physi-
2.1 Compact models for control design cal world), simulation models can include a lot more detail
Models belonging to this class are directly used for con- and second-order effects, since modern CPUs and simula-
troller design, and are usually formulated in state-space tion environments can easily handle complex systems with
form: (tens of) thousands of variables. It is well-known that OOM
ẋ(t) = f (x(t), u(t), p, t) methodologies and EOOLs provide very good support for
(1)
y(t) = g(x(t), u(t), p, t) the development of such models, thanks to equation-based
where x is the vector of state variables, u is the vector of modelling, a-causal physical ports, aggregation and inher-
system inputs (control variables and disturbances), y is the itance. If the OOM model does not contain discrete vari-
vector of system outputs, p is the vector of parameters, and ables and events, then it is basically equivalent to the set of
t is the continuous time. A special case is that of linear, DAEs:
time-invariant models (LTI), which can be described as: F (x(t), ẋ(t), u(t), y(t), p, t) = 0 (5)

ẋ(t) = Ax(t) + Bu(t) Many EOOLs and tools also allow to describe hybrid sys-
(2) tems, with discrete variables, conditional equations or ex-
y(t) = Cx(t) + Du(t)
pressions, and events. For example, see [7, 8] and refer-
or, equivalently, as a transfer function: ences therein for hybrid system descriptions based on hy-
brid automata, or the Modelica language specification [41],
G(s) = C(sI − A)−1 B + D. (3) in particular Appendix C. Although hybrid system control

36
is an interesting and emerging field, for the sake of concise- (5) by giving tentative initial values to the state variables
ness this paper will focus on purely continuous-time phys- (which makes the initialization problem easier to solve)
ical models, with application to the design of continuous- and then to simulate it until it reaches a steady state. If the
time or sampled-time control systems. system is asymptotically stable and the inputs ū are known,
These larger, more detailed models play a double role, this is relatively straightforward; otherwise, it is necessary
with respect to those described in the previous sub-section. to add suitable feedback controllers to drive the outputs to
On one hand they allow to check how good (or crude) the the desired values ȳ and/or to stabilize the system. In both
compact models is, compared to a more detailed descrip- cases, the simulation of this initialization transient might
tion, thus helping to develop good compact models. On fail for numerical reasons before reaching the steady state,
the other hand, they allow to check the closed-loop perfor- due to a bad choice of the initial states.
mance of the controlled system, once a controller design is
available. It is in fact well-known that validating the closed- 3.2 Closed-loop performance assessment by
loop performance using the same simplified model that was simulation
used for control system design is not a sound practice; con- Regardless of the actual design methodology, once the con-
versely, validation performed with a more detailed model is troller has been set up, an OOM tool can be used to run
usually deemed a good indicator of the control system per- closed-loop simulations, including both the plant and the
formance, whenever experimental validation is not possible controller model. Many OOM tools provide model export
for some reason. facilities, which allow to connect an OO plant model with
only causal external connectors (actuator inputs and sensor
3. Overview of current CACSD practice outputs) to a causal controller model in a causal simulation
with EOOLs environment. From a mathematical point of view, this cor-
responds to reformulating (5) in state space form (1), by
As of today, the practising control engineer already gets
means of analytical and/or numerical transformations.
much support from EOOL-based tools for his/her control
system design activities. 3.3 Development of simplified models
3.1 Support to control system synthesis The object-oriented approach, and in particular replaceable
components, allows to define and manage families of mod-
A typical starting point for the design of the control system
els of the same plant with different levels of complexity,
is the analysis of the linearized dynamics of the plant,
by providing more or less detailed implementations of the
around one (or more) steady-state operating conditions.
same abstract interfaces. For example, consider a heat ex-
If the EOOL tool only supports simulation, then one can
changer model: the abstract interface has four fluid connec-
run open-loop simulations of the plant model, subject to
tors, two for the hot fluid inlet and outlet, and two for the
step or to, e.g., pseudo-random binary sequence inputs,
cold fluid inlet and outlet. The corresponding implementa-
and then reconstruct the dynamics by system identification
tion might range from a very simple static model based on
procedures.
log-mean temperatures, with a few algebraic equations, up
A more direct approach, supported by many tools, is to
to a very detailed finite volume model using nonlinear fluid
directly compute the A, B, C, D matrices of the linearized
properties and empirical correlations for heat transfer, and
system around specified equilibrium points, using symbolic
with dozens of state variables and a few hundred algebraic
and/or numerical techniques. The result is usually a high-
equations.
order linear system, which can then be reduced to a low-
This feature of OOM allows to develop simulation mod-
order system by standard techniques for linear model order
els with different degrees of detail (and CPU load) through-
reduction, such as, e.g., balanced truncation.
out the entire life of an engineering project, from prelimi-
A non-trivial issue with both approaches is the compu-
nary design down to commissioning and personnel train-
tation of the equilibrium point (what is sometimes called
ing, within a coherent framework. However, this activity is
DC analysis in the field of electrical circuit simulation). In
based on manual work by the modeller, who needs to de-
a typical setting, the desired steady-state values of the out-
velop the different implementations explicitly. Moreover, it
puts ȳ are known, and the tool must solve the steady-state
is often not easy to obtain compact models such as (1), be-
initialization problem for the system (5):
cause this requires to apply simplifications that may not fit
F (x̄, 0, ū, ȳ, p, 0) = 0 (6) well the abstract component boundaries.

in order to find out the corresponding equilibrium values 3.4 Generation of real-time simulation code
of the inputs ū and of the states x̄. This problem can be An important step in the development of embedded control
numerically challenging, because it often requires solving systems is Hardware-In-the-Loop simulation (HIL), where
large systems of coupled nonlinear equations by iterative the real control hardware is tested by connecting it to a real-
methods, which might fail if the iteration variables are not time simulator, instead of the real plant. Many currently
properly initialized. Currently available OOM tools (and, available EOOL-based tools support automatic generation
in particular, Modelica tools) are still far from providing of efficient real time code starting from fairly large simu-
general robust solutions to this problem. A sub-optimal lation models in the form (5). A common strategy for this
approach to find equilibrium points is to initialize system purpose is to apply inline integration [12, 11] to (5), i.e. to

37
substitute the derivatives with their approximation formu- tions; as a consequence, by providing a representation for
lae (e.g. Euler’s formula), and then solve the system using all the stages of the model manipulation (starting from the
all available numerical and symbolic techniques. translation, going to the flattening, to the model order re-
In order to provide real-time code which is fast enough, duction and so forth) it will be possible to make all the ap-
it is usually important to reduce the model complexity with plications interact at different levels, thus combining posi-
respect to off-line simulation models - this can be done by tive effects from different applications and obtaining better
following the approach sketched in Section 3.3. results.
Exchange formats for model equations and for simula-
3.5 Optimization
tion data should probably be based on the XML language,
Some EOOL and tools support some kind of optimization, for several reasons:
which might be useful for control system design. For ex-
• the tree structure of XML documents easily allows to
ample, the gPROMS language [6] has allowed to declare
represent complex data structures, including symbolic
mixed-integer nonlinear optimization problems for a long
representations of equations;
time. More recently, extensions to the Modelica language
were proposed to formulate optimization problems [2]. • XML documents can be read with standard text-editors
and browsers, thus avoiding all the problem usually
3.6 Future perspectives raised by obscure, ad-hoc binary formats;
It is the authors’ view that EOOL-based tools should sup- • there exists a large base of software (open source and
port advanced control system design problems in a much commercial) for the handling of XML files;
more direct way, by making extensive use of control-
• by re-using this existing software, it is quite straight-
oriented symbolic manipulation techniques. Ideally, it
would be good if the control engineer could develop a de- forward to translate an XML document representing a
tailed simulation model by using object oriented tools and mathematical model into any other equation-based lan-
re-usable model libraries, then automatically obtain sim- guage, and vice-versa;
plified, compact models which are already formulated as • binary XML formats can be used to reduce the verbosity
required by the specific control technique. The availability of XML documents and the cost of parsing them;
of such tools might promote the application of advanced, • there exist some languages (e.g. DTD and XSD) to
model-based techniques that are currently limited by the formally specify the structure of the information the
model development process. XML file must contain.
Being aware that this is a very long-term goal, which
might even require some kind of artificial intelligence, Such standard interfaces for flattened Modelica models
some first steps in this direction are discussed in the fol- and their corresponding simulation data are currently being
lowing sections, with particular reference to the Modelica investigated at Politecnico di Milano using the OpenMod-
language and Modelica compliant tools. elica compiler [16, 1] as a host EOOL environment, and
symbolic manipulation tools such as Mathematica, Maple
4. Basic enabling technologies or Maxima as target environments. If the model is purely
continuous-time, i. e., it is equivalent to the DAE (5), then
The advanced, control-oriented features of future EOOL
MathML [42] on one side, and ModelicaXML [35] on the
tools need some basic enabling technologies and method-
other side might constitute good starting points. If hybrid
ologies to build upon. These are briefly discussed in this
models are considered, one may consider all the languages
section.
developed for the description of hybrid automata in re-
4.1 Open standards for model and data exchange cent years [8], even though the class of hybrid systems
Advanced applications of OOM to control system design which can be described in Modelica with when statements
will most likely require to use different specialized tools is larger than just hybrid automata.
in a coordinated fashion, rather than relying on one-fits- 4.2 Model Order Reduction
all comprehensive software tools. In fact, during the last
decades, the number and the quality of simulation, design Another key enabling technology is represented by mixed
and analysis tools has increased enormously: there is plenty numerical-symbolic Model Order Reduction (MOR) tech-
of open and closed source software for the simulation of niques. These have already been successfully applied to the
physical systems, control synthesis, data analysis, test, val- analysis of electrical circuit models, which are based on
idation, personnel training via a graphical user interface, DAE models such as (5), see [40, 17], and are currently
etc. Some of these tools are developed for specific pur- available in commercial tools such as Analog Insydes [13].
poses, while other are more general in scope (e.g., sym- The MOR strategies are based on the clever application of
bolic manipulation tools, differential equation solvers, data three fundamental steps:
analysis packages). Unfortunately, all this software devel- • specify an allowed error bound, e.g. in terms of per-
opment activity did not follow any standardisation process, centage error of certain steady-state output values cor-
leading to a great diversity in the representation of the in- responding to given constant inputs, or in terms of max-
formation. The definition of standard interfaces will be use- imum deviation of some outputs from a reference tra-
ful for the information exchange between different applica- jectory obtained with given input signals, or in terms

38
 of maximum error of small-signal frequency responses problems. The solution to this simplified problem consti-
around a certain operating point and within a given fre- tutes the first guess for a new problem:
quency interval;
(1 − α)Fe (x̄, 0, ū, ȳ, p, 0) + αFt (x̄, 0, ū, ȳ, p, 0) = 0, (9)
• derive a ranking of all terms in all equations, expressing
how much each term has a significant effect on the which will be solved by varying α from 0 to 1 in small
required modelling accuracy; steps, eventually finding the steady-state solution of system
• remove all terms in ascending order, until the specified (8). In general, this approach should help to reduce (and
hopefully eliminate) the need to manually set initial guess
error bound is reached.
values for iteration variables of initialization problems.
Other MOR techniques exist to reduce large linear sys-
tems, based on concepts such as modal analysis and pro- 5. New functionalities for control system
jection methods; see [38] for a comprehensive overview. design
The application of such MOR tools and techniques, pos-
sibly with extended functionality and algorithms, looks 5.1 Simplified symbolic transfer functions
very promising not only for the simplification of electrical In many interesting cases, the performance of the control
circuit models, but also for the order reduction of generic, system is limited by the dynamic behaviour of the con-
nonlinear DAE models, obtained from the flattening of trolled plant. For example, poorly damped oscillations can
generic EOOL models. This kind of techniques should al- limit the bandwidth of motion control systems, as well as
low to automatically obtain approximated compact models non-minimum phase behaviour. The control engineer can
such as (1), starting from much more detailed simulation gain a lot of useful insight from approximated transfer
models, by formulating specific approximation bounds in functions, where the dependence of the critical dynamic
control-relevant terms (e.g., percentage errors of steady- features from a few physical parameters is clearly visible.
state output values, norm-bounded additive or multiplica- For instance, the natural frequency of a pair of complex
tive errors of weighted transfer functions, or ℓ∞ -norm er- poles in a mechanical system might depend mainly on the
rors of output transients in response to specified input sig- stiffness and on the mass of a certain physical component,
nals). Given the ever-increasing computation power that or, the time constant of a right-half-plan zero in a fluid sys-
can be expected by Moore’s law, the future of these tech- tem might depend on the fluid velocity in a certain point.
niques for CACSD applications definitely appears bright. This is a first case where automatic MOR techniques
could prove extremely useful. Ideally, the user should spec-
4.3 Reliable steady-state initialization and static ify the steady-state operating point, the relevant inputs and
model inversion outputs, and some frequency-weighted error bounds, and
A reliable support to the control engineer’s activity requires get low-order approximated transfer functions of the lin-
to improve the techniques to solve the steady-state equa- earized system, with approximated but explicit dependence
tions (6), which are usually the starting point for any kind of the transfer function features (gains, poles and zeros)
of analysis, including MOR. As pointed out earlier, solv- from the physical model parameters. A suitable combina-
ing (6) requires iterative methods which might fail if not tion of EOOL tools (equipped with model import/export
properly initialized. Troubleshooting can be very frustrat- interfaces) with existing MOR tools like Analog Insydes
ing and time-consuming, and calls for experts of both sim- [13] could provide very interesting results in this direction
ulation methods and domain-specific models. This is not without too much effort.
acceptable in the envisioned framework, which is based on
automatic manipulation by EOOL tools. 5.2 Automatic derivation of LFT models
One option, which is currently being investigated at Po- Once a model has been reduced to a low-order state-space
litecnico, is to introduce extensions to the Modelica lan- form by the combined application of symbolic MOR tech-
guage to support homotopy methods, in a way similar to niques and clever model simplifications as explained in
the approach followed by the SPICE circuit simulation pro- Section 3.3, it might be useful to automatically bring them
gram. The basic idea is that each model has two versions: in the form required for advanced control system design,
the “easy” one, for which it is easier to find a steady- using symbolic manipulation tools. Modern control theory
state solutions, and the “true” version, which is the model provides methods and tools in order to deal with design
to be actually used for simulation. The two models share problems in which stability and performance have to be
the same variables, but use different equations. The system guaranteed also in the presence of model uncertainty, both
model obtained by the aggregation of the “easy” models is for regulation around a specified operating point and for
represented by gain scheduled control system design.
Most of the existing control design literature assumes
Fe (x, ẋ, y, u, p, t) = 0, (7)
that the plant model is given in the form of a Linear Frac-
while the aggregation of the “true” models leads to tional Transformation (LFT) (see, e.g., [46, 27]), a mod-
elling paradigm which is currently an active research topic
Ft (x, ẋ, y, u, p, t) = 0, (8)
in the control engineering and system identification com-
The idea is now to first solve the initialization problem munities. In the robust control framework LFT models con-
for (7), which should not give rise to significant numerical sist of a feedback interconnection between a nominal LTI

39
plant and a (usually norm-bounded) operator which rep- is available. Defining now the deviation variables
resents model uncertainties, e. g., poorly known or time-
varying parameters, nonlinearities, etc. A generic such LFT δx(t) = x(t) − x̄ (12)
interconnection is depicted in Figure 1, where the nominal δu(t) = u(t) − ū (13)
plant is denoted with P and the uncertainty block is de- δy(t) = y(t) − ȳ, (14)
noted with ∆. Note that this representation is extremely
general, and by no means limited to uncertain LTI systems; it is possible to approximate the dynamics of (10) with a
in fact, it is possible to describe any nonlinear DAE system the following linear, parameter-dependent system
by putting all the nonlinear functions in the ∆ block and
by providing an LTI model with direct feedthrough terms ˙
δx(t) = A(p)δx + B(p)δu
, (15)
to describe the algebraic equations. δy(t) = C(p)δx + D(p)δu
LFT models can be used for the design of robust and
gain scheduling controllers, but they can also serve as a where the four matrices A, B, C, D are the Jacobians of the
basis for structured model identification techniques, where two functions f and g:
the uncertain parameters that appear in the feedback blocks ∂f ∂f
are estimated based on input/output data sequences. A(p) = , B(p) = ∂u
∂x
The process of extracting uncertain parameters from the ∂g ∂g
design model of the system to be controlled is a highly C(p) = , D(p) = ∂u .
∂x
complex one. Symbolic techniques play a very important
role in this process: the main use for such techniques is Under suitable assumptions (such as that the state space
to find, via suitable pre-processing steps, equivalent rep- matrices are polynomial or rational functions of the ele-
resentations of rationally dependent parametric matrices, ments of p, see, e.g., [46]) it is possible to transform the
which automatically lead to lower-order LFT representa- system description (15) into an LFT representation (see,
tions. Tools already exist to perform this task [27]. again, Figure 1). As mentioned previously, converting (15)
The LFT modelling problem in its simplest form is as- into an LFT with a ∆ block of minimum dimension is a
sociated with the problem of designing a controller for op- non-trivial symbolic manipulation problem.
eration near a nominal operating point for the system. The An even more challenging formulation of the LFT mod-
problem is then formulated on a local linearised represen- elling problem is the one of simultaneously representing in
tation of the plant to be controlled and is familiarly termed LFT form all the linearisations of interest for control pur-
“pulling out the ∆s”, i.e., it consists of manually or sym- poses of the given nonlinear plant. Indeed, in many control
bolically manipulating the linearised equations in order to engineering applications a single control system must be
separate the nominal part of the plant from the uncertain designed to guarantee the satisfactory closed-loop opera-
one, arranging them in a suitable feedback interconnection. tion of a given plant in many different operating conditions
This reformulation of the plant model lies at the vary basis in the presence of parametric and possibly non parametric
of modern robust control theory and is currently supported uncertainty. The gain scheduling approach to the problem,
by a number of different symbolic manipulation tools. A which has been part of the engineering practice for decades,
recent overview of the state-of-the-art in this research area can be roughly summarised as follows: find one or more
can be found in [18]. As an example, consider a time- scheduling variables α which can completely parametrise
invariant, nonlinear state-space system in the form the operating space of interest (e.g., the flight envelope in
the case of aircraft control) for the system to be controlled;
ẋ(t) = f (x(t), u(t), p)
(10) define a parametric family of linearised models for the plant
y(t) = g(x(t), u(t), p), associated with the set of operating points of interest; fi-
nally, design a parametric controller which can both en-
where p denotes a vector of uncertain parameters, and as- sure the desired control objectives in each operating point
sume that the equilibrium condition x̄, ū, ȳ, which solves and an acceptable behaviour during (slow) transients be-
the steady-state equations tween one operating condition and the other. Many design
techniques are now available for this problem (see, e.g.,
0 = f (x̄, ū, p)
(11) [5, 22, 37]), which can be reliably solved, provided that
ȳ = g(x̄, ū, p)
a suitable model in parameter-dependent form has been de-
rived for the system to be controlled. The goal here would
be to arrive at a representation of the dynamics of the non-
∆ linear system in the form depicted in Figure 2, which is
usually denoted as an LPV-LFT system, where the LPV
u P y acronym stands for Linear Parameter-Varying. The model
structure now includes two feedback interconnections: the
u block ∆(p) takes into account the presence of the uncertain
u parameter vector p, while the block Θ(α) models the effect
Figure 1. Block diagram of the typical LFT interconnec- of the varying operating point, parametrised by the vector
tion adopted in the robust control framework. of time-varying parameters α.

40
 The state-of-the-art of modelling for gain scheduling ical or algorithmic representations of the feedback blocks
can be briefly summarised by defining two classes of mod- Θ(α) and ∆(p). Finally, it is apparent from the short liter-
elling approaches: analytical methods based on the avail- ature review presented above that currently only physical
ability of (relatively) reliable nonlinear equations for the and black-box modelling methods are available, while no
dynamics of the plant, from which suitable control-oriented general purpose CACSD tools capable of combining first
representations can be derived (see, e.g., [28] and the ref- principles models and experimental data in a single control-
erences therein); experimental methods based entirely on oriented model seem to exist. The convergence of the two
identification, i.e., aiming at deriving LPV models for the modelling approaches both in terms of methods and tools
plant directly from input/output data (see, among many would be a very desirable outcome of the research in this
others, [21, 45, 23]). The methods belonging to the first field.
class aim at developing LPV models for the plant to be
controlled by resorting to, broadly speaking, suitable ex- 5.3 Automatic computation of inverse models for
tensions of the familiar notion of linearisation, developed robotic systems
in order to take into account off-equilibrium operation of The design of controllers for non-redundant robotic ma-
the system. As far as experimental methods are concerned, nipulators with N degrees of freedom usually starts from
most LPV identification techniques are based on the as- the equations of motion obtained from the Euler-Lagrange
sumption that the identification procedure can rely on one equations [39]:
global identification experiment in which both the control
input and the scheduling variables are (persistently) excited B(q)q̈ + H(q, q̇)q̇ + g(q) = τ (16)
in a simultaneous way. This assumption may not be a rea- yp = K(q) (17)
sonable one in many applications, in which it would be ∂K
desirable to try and derive a parameter-dependent model yv = q̇, (18)
∂q
on the basis of local experiments only, i.e., experiments in
which the scheduling variable is held constant and only the where q is the N -element vector of Lagrangian coordi-
control input is excited. Such a viewpoint has been consid- nates, which usually correspond to the rotation angles of
ered in [43, 34, 23], where numerical procedures for the the actuator motors, q̇ is the vector of the corresponding
construction of parametric models for gain scheduling on generalized velocities, yp describes the position and ori-
the basis of local experiments and for the interpolation of entation vector of the end effector, yv contains the corre-
local controllers have been proposed. sponding generalized velocities, τ is the vector of general-
To our best knowledge the only documented attempt at ized applied forces corresponding to each degree of free-
deriving control-oriented LFT models automatically from dom (usually the torques applied by rotating actuators),
a nonlinear simulator is presented in [44], where the fo- B(q) is the inertia matrix, H(q, q̇) is the matrix correspond-
cus was on the automatic generation of LFT models for ing to the centripetal, Coriolis, and viscous friction forces,
aerospace applications. Much remains to be done. An while the vector g(q) accounts for the effects of the gravi-
EOOL-based CACSD tool dealing with the generation of tational field; all vectors have dimension N .
control-oriented LFT models should allow to specify some The classical approach to write (16) requires to compute
error bounds for the system approximation (with respect to the so-called direct kinematics (DK), i.e. how the values of
steady-state, transient, and frequency response), the choice q and q̇ translate into the position and motion of the robot’s
of input, output and scheduling variables, and the choice of end effector, then to compute the Lagrange function, i.e. the
parameters to include in the LFT representation. Based on difference between kinetic and potential energy, and apply
that, it should be able to automatically compute the struc- the Euler-Lagrange equations. This can be done manually,
ture of the interconnections defined in Figures 1 and 2 for or using one of the specialized tools available for this task.
the robust and gain-scheduling control design problems, Equations (16)-(18) can then be used as a basis for both
respectively, the state-space matrices of the nominal part P controller design and system simulation.
of the model (either as analytical expressions, if possible, Within an OOM approach, it is possible to save much
or at least as algorithms for their computation) and analyt- time by developing an object-oriented model using an
EOOL, e.g. using the Modelica MultiBody library [33].
Due to the kinematic constraints imposed by the joints, the
original flattened model corresponds to an index-3 DAE,
∆(p)
F (x, ẋ, y, u) = 0, (19)
Θ(α)
which is mathematically equivalent to the Lagrange model
(16)-(18).
u u P y
u Currently available Modelica tools tackle the prob-
u lem by applying specialized algorithms, which exploit the
u knowledge of the topology of the kinematic chain, as well
Figure 2. Block diagram of the typical LFT interconnec- as standard techniques such as BLT partitioning, tearing,
tion adopted in the robust LPV control framework. dummy derivatives and symbolic solution of equations

41
 CT yp q0 v τ
q
.
q IK FC DI R q
yp0 q0 τ .
IK FC R q

u
u Figure 4. Block diagram of inverse dynamics control
Figure 3. Block diagram of computed torque control

the non-linear control problem into a linear, time-invariant

[33]. From a conceptual point of view, a change of state
one. Define a virtual input variable v, which satisfies the
variables x allows to transform (19) into an index-1 system
following equation
F1 (x, ẋ, y, u) = 0, (20)
τ = B(q)v + H(q, q̇)q̇ + g(q). (26)
where
      Since the inertia matrix B is structurally non-singular, it is
xp q yp
x= = , y= , u = τ. (21) always possible to solve (26) for v:
xv q̇ yv

Eventually, efficient procedures are produced to solve (20) v = B −1 (q) (τ − H(q, q̇)q̇ − g(q)) . (27)
for ẋ and y given x and u, thus actually bringing the system
into state-space form: Plugging v in the robot dynamics equation (16), one ob-
tains:
ẋ = f (x, u)
(22) q̈ = v. (28)
y = g(x, u).

This formulation can be used to solve simulation problems,

The block diagram interpretation of these equations is
by linking it to any ODE/DAE solver. However, there are
shown in Fig 4: thanks to the dynamic inversion (DI) block,
several other interesting things that could be done with
the dynamic relationship between the virtual input v and
(20), from a control engineer’s perspective.
the Lagrangian positions and velocities q and q̇ (repre-
Robot trajectories are originally defined in terms of
sented by the dotted block) is now described by a simple
end effector coordinates as functions of time yp0 (t). In or-
integrator and a double integrator, respectively. It is then
der to obtain the corresponding reference trajectories in
easy to tune a fixed-parameter, linear feedback controller
Lagrangian coordinates for the low-level robot joint con-
(FC) in order to obtain the desired closed-loop dynamics.
trollers, (17)-(18) must be solved for q, q̇, thus computing
Starting from the index-1 DAE robot model (20), it is
the so-called inverse kinematics (IK):
straightforward to derive the equations and then the explicit
q0 = K −1 (yp0 ) (23) algorithms to compute the DK, IK, CT, and DI, by using the
 −1 same techniques employed to bring (20) into state-space
∂K form. The DK (17)-(18) is obtained by solving (20) for yp
q̇ 0 = yv0 ; (24)
∂q (and possibly yv ) given q (and possibly q̇), while the IK is
obtained by solving (20) for q (and possibly q̇) given yp
note that the Jacobian of K(q) is also needed to solve
(and possibly yv ); the subset of required equations is found
(23), since analytical inverses cannot usually be obtained.
by suitable analysis of the incidence matrix. The CT (25) is
Furthermore, two interesting approaches to model-based
obtained by solving (20) for τ given q, q̇, and q̈. Finally,
robot control are based on suitable manipulations of eq.
the DI (26) is obtained by solving (20) augmented with
(16): the pre-computed torque approach and the inverse
(26) for τ given v, q, and q̇. EOOL tools should then be
dynamics approach [39].
able to automatically generate the code corresponding to
The pre-computed torque approach is a feed-forward
the DK, IK, CT, and DI blocks in two forms: as algorithms
compensation scheme, where the theoretical torque re-
to compute the outputs given the inputs (e.g., C code for
quired to follow the reference trajectory is directly fed to
direct inclusion in the robot controller), as well as equation-
the torque actuators (see Fig. 3) in order to obtain a good
based Modelica blocks, which could be used for closed-
dynamic response to the set point yp0 . The CT block per-
loop simulation within a Modelica environment.
forms this task by solving (16) for τ , given the reference
As a final remark, note that the method of inverse dy-
trajectory and its derivatives:
namics is a special case of the much more general theory
τ = B(q 0 )q̈ 0 + H(q 0 , q̇ 0 )q̇ 0 + g(q 0 ). (25) of feedback linearization [20], whose goal is to obtain a LTI
dynamics made by pure integrators from generic nonlinear
A feedback controller (FC) is also included to deal with systems, by applying suitable feedback actions as shown
uncertainties and disturbances. in Figure 4. It could also be interesting to investigate the
The inverse dynamics approach is a feedback compen- coupling between EOOL tools and symbolic manipulation
sation scheme, that uses the model in order to transform tools for the design of such controllers.

42
5.4 Fast and compact models for Model Predictive The main requirement for NMPC-oriented models is
Control that they must have the least possible number of state and
The Model Predictive Control (MPC) approach [25, 36] is algebraic variables, in order to keep the complexity of the
based on a few key ideas, that turn the control problem into optimization problem within acceptable limits, and that
an optimization problem. The control variable is a discrete- they have good smoothness properties, in order to avoid
time variable, that changes periodically every Ts seconds: convergence problems of the iterative optimization algo-
rithms. The development of those models can be very time
u(t) = u(k), kTs ≤ t < (k + 1)Ts . (29) consuming, and require highly skilled manpower; it is ap-
parent how better tool support could be extremely useful in
At each time step k, an optimization problem is solved, order to reduce the development effort and cost.
whose unknowns are the next values of the control variable The potential of OOM for MPC was first noted by Ma-
u(k + i) over a finite horizon 1 ≤ i ≤ N . The first sample ciejowski at the end of the ’90 [24]. There are several re-
u(k + 1) is then applied to the actuators at the next time ported case studies [14, 15, 3, 19], where the model used in
step, the rest of the values are discarded, and the process the NMPC algorithm was derived from a Modelica model
is repeated over and over, thus implementing a receding of the physical plant, using the tool Dymola to produce
horizon strategy. the code corresponding to the state-space form (1), i.e., the
There are different ways to formulate the MPC prob- dsmodel.c code that is usually linked to ODE/DAE solvers.
lem, depending on the specific technique used to solve the In order to derive suitably simplified models, the features
problem. Generally speaking, the figure of merit to be min- of Modelica discussed in Section 3.3 have been extensively
imized is a quadratic function, which suitably weights the exploited. In general, this approach has proven much more
future deviations of the controlled variables from the set satisfactory than writing the C-code of the model from
point and the intensity of the control action, as well as any scratch; however, it still requires a substantial investment
other problem-specific performance index that has to be of time and effort for each new application.
minimized, e.g. the financial cost of running the process. The application of the automatic MOR techniques de-
The constraints of the optimization problem are the dy- scribed in section 4, possibly still combined with some
namic relationship between the input and output variables, manual intervention in terms of replaceable models, looks
typically in the form (4), and possibly other constraints, very promising in order to bring detailed simulation mod-
such as upper and lower bounds of the state, control and els into a form which is suitable for NMPC with a much
output variables and of their rate of change. more limited effort by the developer.
The main advantage of MPC is its intrinsic ability to Furthermore, [19] correctly points out that, although
deal with highly interacting multivariable systems (many the interface to NMPC algorithms is very similar to the
control inputs and controlled outputs), while keeping into interface to ODE/DAE solvers, the former requires some
account operating constraints such as actuator saturations more flexibility. For example, advanced NMPC schemes
or hard bounds on controlled variables, and at the same can provide on-line estimation of uncertain parameters
time meeting some problem-specific optimality criterion. through the use of extended or unscented Kalman filters.
The main drawback is the high computational load, since This means that some model parameters are no longer con-
a (possibly non-linear and non-convex) constrained opti- stant throughout a transient, so that the C-code obtained
mization problem must be solved at each sampling time; for simulation purposes must be manually adapted. A bet-
this makes MPC suitable for systems with slow dynamics, ter option would be to implement a code export interface
e. g. chemical plants, where there is plenty of time to carry which makes it possible to turn selected parameters ap-
out the required computations in real time. This limitation pearing in (5) (which are going to be estimated on-line)
is likely to become less and less stringent in the future, into inputs, before transforming the system in state-space
thanks to Moore’s law. form (1).
The second issue is the requirement that a suitable plant
model is available, as the control system performance criti-
cally depends on the model quality. Models for linear MPC
can be obtained either by linearization of analytical models,
or by system identification from experimental and/or simu- 6. Conclusions
lation data, e. g. step responses; both cases are already sup- After a brief review of the different uses of models in con-
ported by current EOOL tools. Nonlinear MPC (NMPC) trol system design, the current state of the art of EOOL-
algorithms are preferably based upon analytical models in based tools for CACSD has been reviewed: apparently, cur-
state-space form (1), which are derived from physical first- rently available tools mainly focus on simulation tasks.
principles models. The conversion to discrete-time form (4) Several further directions for research and development
is often performed internally by the NMPC algorithm it- in EOOL tools where then discussed, which go beyond
self, by standard ODE integration routines. This means that the mere simulation problem. Results in these directions
the interface between the EOOL tool and the NMPC tool could substantially improve the level of support to the con-
is similar to the one used for simulation problems, i.e. the trol engineer willing to apply advanced, model-based con-
state-space form (1), possibly augmented by the Jacobians trol techniques to real-life problems, starting from object-
of the right-hand-sides of (1). oriented models of the plant.

43
References [15] R. Franke, M. Rode, and K. Krüger. On-line optimization
[1] OpenModelica home page. URL: http://www.ida. of drum boiler startup. In Proceedings 3rd International
liu.se/labs/pelab/modelica/OpenModelica. Modelica Conference, pages 287–296, Nov. 3–4 2003.
html. [16] P. Fritzson, P. Aronsson, A. Pop, H. Lundvall, K. Nyström,
[2] J. Åkesson. Optimica - An extension of Modelica L. Saldamli, D. Broman, and A. Sandholm. OpenModelica -
supporting dynamic optimization. In Proceedings 6th A free open-source environment for system modeling, sim-
International Modelica Conference, pages 57–66, Bielefeld, ulation, and teaching. In Proceedings IEEE International
Germany, Mar. 3–4 2008. Symposium on Computer-Aided Control Systems Design,
Munich, Germany, Oct. 4–6 2006.
[3] J. Åkesson and O. Slätteke. Modeling, calibration and
control of a paper machine dryer section. In Proceedings [17] T. Halfmann and T. Wichmann. Symbolic methods in
5th International Modelica Conference, pages 411–420, industrial analog circuit design. In Scientific Computing in
Vienna, Austria, Sep. 4–5 2006. Electrical Engineering, Mathematics in Industry. Springer
Verlag, 2006.
[4] M. Andersson, S. E. Mattsson, D. Brück, and T. Schöntal.
OmSim - an integrated environment for object-oriented [18] S. Hecker and A. Varga. Symbolic manipulation techniques
modelling and simulation. In Proceedings of the IEEE/IFAC for low order LFT-based parametric uncertainty modelling.
Joint Symposium on Computer-Aided Control System International Journal of Control, 79(11):1485–1494, 2006.
Design, CACSD’94, pages 285–290, Tucson, Arizona, [19] L. Imsland, P. Kittilsen, and T. S. Schei. Model-based
March 1994. optimizing control and estimation using modelica models.
[5] P. Apkarian and R. J. Adams. Advanced Gain-Scheduling In Proceedings 6th International Modelica Conference,
Techniques for Uncertain Systems. IEEE Transactions on pages 301–310, Bielefeld, Germany, Mar. 3–4 2008.
Control System Technology, 6:21–32, 1998. [20] H. K. Khalil. Nonlinear Systems. Prentice Hall, 3rd edition,
[6] P. I. Barton and C. C. Pantelides. The modeling of combined 2002.
continuous and discrete processes. AIChE Journal, 40:966– [21] L. Lee and K. Poolla. Identification of linear parameter-
979, 1994. varying systems using nonlinear programming. Journal of
[7] D.A. van Beek, M.A. Reniers, J.E. Rooda, and R.R.H. Dynamic Systems, Measurement and Control - Transactions
Schiffelers. Foundations of an interchange format for hybrid of the ASME, 121(1):71–78, 1999.
systems. In A. Bemporad, A. Bicchi, and G. Butazzo, [22] D. J. Leith and W. E. Leithead. Survey of gain-scheduling
editors, Computation and Control, 10th International analysis and design. International Journal of Control,
Workshop, volume 4416 of Lecture Notes in Computer 73(11):1001–1025, 2000.
Science, pages 587–600. Springer Verlag, 2007. [23] M. Lovera and G. Mercere. Identification for gain-
[8] D.A. van Beek, M.A. Reniers, J.E. Rooda, and R.R.H. scheduling: a balanced subspace approach. In 2007
Schiffelers. Concrete syntax and semantics of the composi- American Control Conference, New York, USA, 2007.
tional interchange format for hybrid systems. In Proc. 17th [24] J. M. Maciejowski. Modelling and predictive control:
IFAC World Congress, Seoul, Korea, Jul 6–11 2008. enabling technologies for reconfiguration. Annual Reviews
[9] E. Carpanzano and C. Maffezzoni. Symbolic manipulation in Control, 23(1):13–23, 1999.
techniques for model simplification in object-oriented [25] J. M. Maciejowski. Predictive control: with constraints.
modelling of large scale continuous systems. Mathematics Prentice Hall, 2002.
and Computers in Simulation, 48(2):133–150, 1998.
[26] C. Maffezzoni and R. Girelli. Modular modelling in
[10] F. E. Cellier and H. Elmqvist. Automated formula an object-oriented database. Mathematical Modelling of
manipulation supports object-oriented continuous-system Systems, 4:121–147, 1998.
modeling. IEEE Control Systems Magazine, 13(2):28–38,
1993. [27] J.-F. Magni. Linear fractional representation toolbox.
Technical Report TR 6/08162 DCSD, ONERA, 2004.
[11] H. Elmqvist, S. E. Mattsson, and H. Olsson. New methods
for hardware-in-the-loop simulation of stiff models. In [28] A. Marcos and G. Balas. Development of linear-parameter-
Proceedings 2nd International Modelica Conference, pages varying models for aircraft. Journal of Guidance, Control
59–64, Oberpfaffenhofen, Germany, Mar. 18–19 2002. and Dynamics, 27(2):218–228, 2004.

[12] H. Elmqvist, M. Otter, and F. Cellier. Inline integration: [29] S. E. Mattsson, M. Andersson, and K. J. Åström. Modeling
A new mixed symbolic /numeric approach for solving and simulation of behavioral systems. In Proceedings of the
differential–algebraic equation systems. In Proc. ESM’95, 32nd IEEE Conference on Decision and Control, volume 4,
European Simulation Multiconference, pages xxiii–xxxiv, pages 3636–3641, San Antonio, Texas, Dec. 1993.
Prague, Czech Republic, Jun. 5–8 1995. [30] S. E. Mattsson, M. Andersson, and K. J. Åström. Object-
[13] The Fraunhofer-Institut für Techno-und Wirtschafts- oriented modelling and simulation. In D. Linkens, editor,
mathematik. Analog Insydes. URL: http://www. CAD for Control Systems, pages 31–69. Marcel Dekker
analog-insydes.de/. Inc., New York, 1993.

[14] R. Franke. Formulation of dynamic optimization problems [31] S. E. Mattsson and H. Elmqvist. Simulator for dynamical
using modelica and their efficient solution. In Proceedings systems using graphics and equations for modeling. IEEE
2nd International Modelica Conference, pages 315–323, Control Systems Magazine, 9(1):53–58, Jan. 1989.
Oberpfaffenhofen, Germany, Mar. 18-19 2002. [32] S. E. Mattsson, H. Elmqvist, and M. Otter. Physical system

44
 modeling with Modelica. Control Engineering Practice,
6(4):501–510, 1998.
[33] M. Otter, H. Elmqvist, and S. E. Mattsson. The new Mod-
elica MultiBody library. In Proceedings 3rd International
Modelica Conference, pages 311–330, Linköping, Sweden,
Nov. 3–4 2003. URL: http://www.modelica.org/
events/Conference2003/papers/h37_Otter_
multibody.pdf.
[34] B. Paijmans, W. Symens, H. Van Brussel, and J. Swevers. A
gain-scheduling-control technique for mechatronic systems
with position-dependent dynamics. In Proceedings of the
2006 American Control Conference, Minneapolis, USA,
2006.
[35] A. Pop and P. Fritzson. ModelicaXML: A Modelica XML
representation with applications. In Proceedings of the
3rd International Modelica Conference, pages 419–430,
Linköping, Nov 3–4 2003.
[36] S. J. Qin and T. A. Badgwell. A survey of industrial
model predictive control technology. Control Engineering
Practice, 11:733–764, 2003.
[37] W. Rugh and J. Shamma. Research on gain scheduling.
Automatica, 36(10):1401–1425, 2000.
[38] P. Schwarz, J. Bastians, C. Clauss, J. Haase, A. Köhler,
G. Otte, and P. Schneider. A tool-box approach to computer-
aided generation of reduced-order models. In Proceedings
EUROSIM 2007, 2007.
[39] L. Sciavicco and B. Siciliano. Modelling and Control of
Robot Manipulators. Springer Verlag, 2000.
[40] R. Sommer, T. Halfmann, and J. Broz. Automated
behavioral modeling and analytical model-order reduction
by application of symbolic circuit analysis for multi-
physical systems. In Proceedings of EUROSIM 2007,
Ljubljana, Slovenia, Sep 9–13 2007.
[41] The Modelica Association. Modelica - A unified object-
oriented language for physical systems modeling - Lan-
guage specification version 3.0. Online, Sep. 5 2007.
URL: http://www.modelica.org/news_items/
documents/ModelicaSpec30.pdf.
[42] The World Wide Web Consortium. Mathematical Markup
Language (MathML) Version 2.0. Online, Oct 21 2003.
URL: http://www.w3.org/TR/MathML2/.
[43] J.J.M. van Helvoort, M. Steinbuch, P.F. Lambrechts, and
R. van de Molengraft. Analytical and experimental
modelling for gain-scheduling of a double scara robot.
In Proceedings of the 3rd IFAC Symposium on Mechatronic
Systems, Sydney, Australia, 2004.
[44] A. Varga, G. Looye, D. Moormann, and G. Gräbel.
Automated generation of LFT-based parametric uncertainty
descriptions from generic aircraft models. Mathematical
and Computer Modelling of Dynamical Systems, 4(4):249–
274, 1988.
[45] V. Verdult. Nonlinear system identification: a state space
approach. PhD thesis, University of Twente, 2002.
[46] K. Zhou, J. U. Doyle, and K. Glover. Robust and optimal
control. Prentice-Hall, New Jersey, 1996.

45