Professional Documents
Culture Documents
Amir Noori 2007 HYBRID SYSTEMS MODELING FOR GAS TRANSMISSION NETWORK
Amir Noori 2007 HYBRID SYSTEMS MODELING FOR GAS TRANSMISSION NETWORK
Amir Noori 2007 HYBRID SYSTEMS MODELING FOR GAS TRANSMISSION NETWORK
The 4th International Federation of Automatic Control Conference on Management and Control of Production and Logistics
September 27-30, Sibiu - Romania
389
restrict the behavior of a plant such that the given controller. However, the membership degree of this
specifications are fulfilled as much as possible (for variables are as our control variables. In a decision
further reading, see (Ramadge, et al., 1989)). Today, support system, which can be used in supervisory
this approach is common in discrete event system control systems, these transitions yield using if-then
modeling and analysis and many researches are rules (and some other fusion rules).
employing this framework. But, RW framework is The major advantage of our model is its interactivity
based on automata and language theory (and and reconfigurability, which makes it a suitable
symbolic computation), which is far from linear framework for diagnosing and management purpose.
system theory principles. Also, based on this model, analysis and systematic
Moreover, for analysis of these types of problems, design of control system is possible.
which often meet in lower level analysis of hybrid The rest of the paper is organized as follows. In
systems many approaches have been proposed section 2, we present a brief introduction to Natural
(Cohen, et al., 1989; Murata, 1989; Harel, 1987; Ho, Gas Transmission Network, followed by section 3
1989). In linear system theory, we have many that models this network based on hybrid automaton.
powerful tools for analysis of dynamical system, In section 4, This network is then implemented in
which can reuse in discrete event systems but most SIMULINK TM /Stateflow TM which, provides a suitable
of mentioned approach does not have strong test bed for further researches. Section 5 is devoted to
connection to classical system theory. present control system design. Section 6 concludes
In 1981, the researchers in INRIA started working on the paper.
what we now know as Max-algebra; a system
theoretic approach for discrete event systems.
Although, this approach is well suited in some 2. NATURAL GAS TRANSMISSION NETWORK
applications (Doustmohammadi, 1993; Kamen,
1993) but after twenty-six years, it is not as general Natural gas pipeline networks, or pipeline systems,
as linear system theory. Recently, fuzzy discrete are used to transport gas from sources to consumers
event systems (FDES) are introduced which convert over long distances. The distance from a source to a
symbolic computation of DES to a mathematical consumer may be thousands of kilometers. The gas
framework. in the pipeline is pressurized in order to maintain a
In this paper, we present a hybrid model of gas pressure difference necessary for moving the gas.
transmission network. This network can model using Compressors are used to pressurize the gas. These
hybrid automaton. Hybrid automaton is a suitable are needed at regular intervals along the pipeline-
tool for modeling large-scale systems. Moreover, in usually every 50 to 150 kilometers, since the gas
contrast with petri-nets and its extensions, this pressure decreases rapidly due to frictional losses.
framework brings decidability property, which in The most elementary components of a pipeline
complex systems is an important subject. A problem system are pipeline segments and compressors. Other
is said to be decidable if there exists an algorithm to components of pipeline system are valves (discrete or
solve that problem and otherwise, is said to be continuously operating) and gas storages. Long-
undecidable. distance transmission pipelines operate under high
However, utilization of this feature needs (or pressures and gas users at the off-takes use pressure
perhaps, are easier) converting the underlying hybrid reduction stations to adapt the gas pressure to their
system to a unified discrete event model. Then, we needs (Alato, 2005).
can decide about important properties of this discrete However, in this paper we only investigate the
event system and extend these results to the concrete following components: Pipeline Segments,
one. Compressors (compressor Stations), Line Break
Today, the analysis of complex automation systems Valves and other type of valves. Modeling of these
becomes more important. However, dependencies components is done based on hybrid automaton. A
between infrastructures such as power transmission major advantage of this modeling approach is its user
network, oil and gas transmission network and friendliness that can be easily used and generalized
telecommunication network make this issue for other relevant applications. In pipeline systems,
prominent. In reality, we have some events that are compressor stations are the most complex parts of the
not observable to operator, because of the lack of model. Fig.1, illustrates connections between
sensors or detection means, too high cost of components in a compressor station.
detection, or difficulty in information transmission. Connection Valve
Compressor
Pipeline
Thus, some behavior of the underlying system is not Line1
observable to user.
In this paper, we first model some important Line2
390
Where k12 is a constant and z12 is the average
3. HYBRID MODELING OF COMPONENTS
compressibility of the gas in the segment. If the
elevation of the pipeline segment is essential, then
In this section, we provide a typical hybrid model for
the following expression yields:
each component of gas transmission network. As
mentioned above, we have pipeline segments, k12 g (h2 − h1 ) p122
p12 − p 22 − = k12 z12 q12 q12
0.8539
compressor stations and various types of valves as z12T12
major components of gas transmission network.
The following definition is based on (Johansson, et (2)
al., 1999; Lygeros, et al., 1999).
Where
Definition 1 (Hybrid Automaton). A hybrid h2 is the elevation of the end of the pipeline segment
automaton H is a 6-tuple H = (Q, X, X0 , f, D, R) , above some basic level
Where h1 is the elevation of the beginning of the segment
Q is a finite set of discrete variables; above some basic level
X is a finite set of continuous variables with T12 is the average gas temperature of the segment
X 0 ⊆ Q × X is a set of initial states;
The average pressure p12 in the segment is
f : Q × X → TX is a vector field; calculated using the expression:
D ⊆ Q × X is the domain of H; 2⎛ p p ⎞
R : Q × X → P (Q × X ) is a reset relation. p12 = ⎜⎜ p1 + p 2 − 1 2 ⎟⎟ (3)
3⎝ p1 + p 2 ⎠
We refer to (q, x ) ∈ Q × X as the state of H. In the leak mode, we assume the pressure of nodes
decreases exponentially. In the break mode, pressures
The MathWorks modeling and simulation tools, assumed to be zero. a hybrid model of a pipe
Simulink™/Stateflow™ facilitate the design of segment is shown in fig.2.
complex control systems. In the following, we also Fig.3 shows implementation of a pipe segment in
describe modeling of these components in Stateflow TM .
Simulink™ environment.
391
subsystem in SIMULINK TM environment. In Fig.4,
this subsystem is shown.
In transient condition, based on the identification As we can see in equation (4), the output pressure of
methods and acquired data from simulation of compressor has a nonlinear algebraic relation with
network in Simone TM , we obtained a transfer function the input pressure. To represent such behavior for
compressor, we use a subsystem that invokes through
for pipe segments with an acceptable error, which is
a function call in Stateflow TM . This subsystem is
useful for optimization and control purpose. But,
these linear models of components are very sensitive shown in Fig.6.
to component’s parameters and individually should
be computed for each component of network.
Therefore, in this paper, we will not discuss this
subject.
3.2 Compressors
392
However, the fail state rarely uses but it is useful for
diagnosing and management purpose.
Plant .
(Hybrid model)
393
However, defining these levels depend on plant and
desired operation features. In small-scale plants, the
second level is unnecessary and user can simply
interact with the concrete system. In Fig.10, it is
obvious that the abstraction process reduces the
knowledge and decision-making takes place under
incomplete knowledge.
However, in large-scale systems (such as gas
transmission network) interactions between operator
and the underlying process are very complex and
needs a systematic approach for analysing the
complex behaviours.
In this paper, we concentrate on the level two and
propose an approach for designing an automated Fig.11 Fuzzy abstraction of hybrid automaton
control system (ACS). We can present this level by
Discrete Event Systems (DESs) (or Fuzzy DES). Alternatively, we can formulate problem in different
Designing controller (which is named supervisor) for manner to solve it. The supervisor gets positive
such systems widely considered in literatures and has reinforcement for maintaining pressure near the
many application. desired value, negative reinforcement for large
One way for designing an ACS is by using Expert deviations from this value. Then, supervisor can
systems and Decision Support Systems (DSSs). learn the best strategy, to operate network by
However, some authors proposed intelligent providing different setpoints for station and
approach, which effectively handles uncertainties in producers, and trying to maximize the reinforcement,
knowledge and model (Nokhbeh, et al., 2006). This it receives. This strategy is easy to implement and
task is accomplished using fuzzy variables, which scalable also. This way of solving problem is know
generally has gaussian membership function. as RL (Sutton, et al., 1998).
However, implementations of such systems are In the follownig, we discuss a learning method for
difficult and are not scalable. tunning variables of supervisor automaton.
In the following, we present a new approch, which
are based on a hybrid model of network and deals
with model and knowledge uncertainties in a 5.3 Reinforcement Learning
computational framework.
Machine learning is study of algorithms, which
improve performance with experience
5.2 Fuzzy Abstraction of Hybrid model (Mitchell,1997). Such an algorithm is known as
learner and procedure of improving performance
Reduction or abstraction techniques replace a with experience is known as training experience.
complex system with a more abstract one, and partly RL is learning from interaction with environment.
soften the state explosion problem. There are two The learner and decision-maker is called the agent.
common method for this purpose; predicate The thing it interacts with, is called the environment.
abstraction which divide state space into several These interact continually, the agent selecting actions
distinct regions, and projection which simply remove and the environment responding to those actions and
some less important variables. presenting new situations (known as states) and
By considering uncertainties, which naturally exists reward associated with them to the agent, as shown
in the concrete model, we extend first approach by in Fig.12. There are two flavors of RL problem based
dividing state-space into fuzzy subsets. The resulting on the fact, that the agent is aware of the
model is a Fuzzy DES that can be represented in environmental model or not. These are known as
matrix form. In fig.11, State diagram of a fuzzy model-based and model-free reinforcement learning.
automaton is shown. Transistions in this diagram are Q-learning is an example of model-free RL and
carried out using a matrix which represent possibility pritorized sweeping is a model-based RL. In model-
of each transition. In supervisory control system, this based approaches, we learn a model and based on the
variables are as our control variable. estimated model make a decision. In model-free RL,
However, transitions of such systems are somewhat no effort are used to estimate a model.
vague, and can be represented by a variable structure
automaton. A variable structure automaton is an Agent
automaton that their transitions can tune. Thus, using State(Si) Action(Ai)
these abstracted model, one can design a supervisor
based on if-then rules. In other words, one way to Reward(Ri)
solve this problem is to manually code rules for
controling subsystems and their interactions. Environment
Nevertheless, as previously mentioned, this approach
is complex to implement and not easy to scale.
Fig.12 RL model
394
a11 : Open_valve_11
Watkins' Q-learning, is a model-free method which a12 : Open_valve_13
is very easy to implement . Q-learning works by
estimating the value of state-action pair, Q(s; a). The We also used a simple cost function for this purpose,
value Q(s; a), known as Q-value, is defined to be the which give all needed information about these two
expected sum of future reinforcement (penalty) states :
obtained by taking action a from state s and
following an optimal policy thereafter. Once these CostFun = (press_1_1–500)^2 + (press_1_2– 500)^2
values have been learned, the optimal action from + (press_1_3–500)^2 + (Flwrate_1_1 – 120)^2 +
any state is the one with the lowest Q-value.The (Flwrate_1_2–120)^2 + (Flwrate_1_3–120)^2
estimation of Q-values can be done on the basis of (6)
experience using following learning rule (Sutton, et Finally, after the training the supervisor, Action 3 in
al., 1998): state1 and action1 in state4 is selected as best actions
Q( s, a ) = Q( s, a ) + α ( X + min (Q( s′, a′) − Q( s, a ))) with minimum q-values.
a′∈A Fig.15, shows that the action with lowest Cost
function, is selected more than others and fig.16
(5) represent the convergence of Q-values. However, in
where, this case, there are one Optimal Policy as mentioned
s0 is new state after taking action a on state s above, First, Open_valve_7, then Open_valve_5.
and X is reinforcement observed.
X = PathCost , if transition is safe
X = R , if transition is unsafe
α is known as learning rate.
A part of IGTN is shown in fig.13 which is used for
explainig the implementation of intelligent control
system.
6. CONCLUSION
395
7. REFERENCES Conference on Decision and Control, Phoenix,
AZ.
K. S. Chapman, M. Abbaspour, (june,2002) H. J. Sandler and E. T. Luckiewicz, , (1987)
“Virtual Pipeline System Testbed to Optimize “Practical Process Engineering”, McGraw-Hill,
the U.S. Natural Gas Transmission Pipeline 638 pages.
System”, Technology Status Assessment Report, H. Nokhbeh, M. B. Menhaj, (2006) “fuzzy decision
The Department of Energy, Strategic Center for support system for crisis management in gas
Natural Gas, Kansas State University. transmission network”, Proceedings of the 15th
H. Alato, (2005) “Real-time Receding Horizon International Iranian Conference on Electrical
Optimisation of Gas Pipeline Networks” PHD Engineering.
dissertation, Helsinki University of Technology. N.R. Jennings, P. Faratin, A.R. Lomuscio, S.
available on-line : Parsons, C. Sierra, M. Wooldridge, (2001)
http://lib.tkk.fi/Diss/2005/isbn9512276593 “Automated negotiation: Prospects, methods and
J. Osiadacz, S. Swierczwski, (1994) “Optimal challenges”, Internat. J. Group Decision
Control of Gas Transportation Systems”. The Negotiation 10 (2) 199–215.
3rd IEEE Conference on Control Applications, T. M. Mitchell, (1997) “Machine Learning”.
Yhe University of Strathclyde, Glasgow. McGraw Hill.
R. G. Carter, “Pipeline Optimisation: (1998) R. S. Sutton and A. G. Barto, (1998) “Reinforcement
“Dynamic Programming after 30 years”, Learning: An Introduction”. London, England:
Pipeline Simulation Interest Group, Annual The MIT Press.
Conference ,Denver, Colorado, F. Lin, H. Ying, (2002) “Modeling and Control of
USA,www.psig.org/papers Fuzzy Discrete Event Systems” IEEE
Kelling C., K. Reith and E. Sekirnjak (2000), “A TRANSACTIONS ON SYSTEMS, MAN, AND
Practical Approach to Transient Optimisation for CYBERNETICS—PART B: CYBERNETICS,
Gas Networks”, Pipeline Simulation Interest VOL. 32, NO. 4
Group, Annual Conference , Savannah, C.G. Cassandras and S. Lafortune, (1999)
Georgia, USA, www.psig.org/papers “Introduction to Discrete Event Systems”,
X. D. Koutsoukos, P. J. Antsaklis, J. A. Stiver, Kluwer Academic Publishers.
AND M. D. Lemmon, (2000) “Supervisory “MATLAB&SIMULINK user’s guide”, The
Control of Hybrid Systems”, Proceedings of the MathWorks, Inc (2005).
IEEE, VOL. 88, NO. 7 A. Noori, M. B. Menhaj, (2007) “Fuzzy abstraction
P.J.G. Ramadge andW.M.Wonham. (1989) “The of hybrid systems” , will appear in : Proceedings
control of discrete event systems”. IEEE of International conference on application of
Proceedings: Special issue on Discrete Event Computational Intelligence for Measurement
Systems, 77(1):81–97,. Systems and Applications.
G. Cohen, P. Moller, J.-P. Quadrat and M. Viot,
(1989) “Algebraic Tools for the Performance
Evaluation of Discrete Event Systems”, Proc.
IEEE, Vol 77, No 1,
T. Murata, (1989) “Petri nets: Properties, analysis
and applications”, Proceedings of the IEEE. Vol
77 No. 4.
D. Harel, (1987) “Statecharts: A Visual Formalism
for Complex Systems”, Science of Computer
Programming, vol. 8.
Y.-C. Ho, (1989) “Special Issue on the Dynamics of
Discrete Event Systems”, Proceedings of the
IEEE.
Doustmohammadi, (1993) “Modeling of Discrete
Event Dynamic Systems based on Petri Nets and
Minimax Algebra Approach”, PhD thesis
proposal, Georgia Inst. Tech., Atlanta.
E. W. Kamen, (1993) “An Equation-Based Approach
to the Control of Discrete Even Systems with
Applications to Manufacturing”, Proc. Int’l
Conf. on Control Theory & its Applications,
Jerusalem, Israel.
K. H. Johansson, M. Egerstedt, J. Lygeros, and S.
Sastry. (1999)”On the regularization of Zeno
hybrid automata.” Systems & Control Letters,
38:141-150.
J. Lygeros, K. H. Johansson, S. Sastry, and M.
Egerstedt. (1999) “On the existence of
executions of hybrid automata.” In IEEE
396