Professional Documents
Culture Documents
Distributed: Intelligent Control For Mine Refrigeration System
Distributed: Intelligent Control For Mine Refrigeration System
Surface Dam
Underground Dam
$4 Control Valve
~~
Fig. 3. Exumple of rulesfor the underground agent plan is based on the refngeration sys-
tem conditions and state known to the
in the distributed controller, the SCADA system can be consid- agent, e.g., the underground agent checks heater availability,un-
ered as simply another agent. derground dam level, heater inlet temperature, etc. If the up-
stream agent cannot satisfy the plan locally’, it returns an
Distributed Controller intermediate plan to the downstream agent and it forwards a sec-
The refrigeration system is intended to be operated as fol- ondary plan to its own upstream agent. This upstream agent tries
lows. A sequence of demand intervals for heated (cooled) water to satisfy the proposed secondary plan.
is specified by the system operator. Each demand interval com- An example illustrating the typical planning process is pro-
prises a start and stop time and the required temperature and vided in Fig. 4. Obviously, the details of the plans passed be-
flow for that time interval. It is assumed that the demand is tween the agents depend on actual demand values and plant
specified well in advance of when it is actually required in order conditions. For example, if due to maintenance heaters are un-
to avoid difficulties due to unreasonable deadlines. For exam- available, the intermediate plans are changed to reflect this, and
ple, at 08h00 a demand of 40 degrees C at 2 I/min for 12h00- different reasoning is applied to develop the final plan. As is evi-
16h00, and then 35 degrees C at 4 l/min for 20h00-Oh00 with no dent, the plan is modified at each stage of the planning process,
demand from 16h00-201100.Details of a distributed intelligent e.g., if an agent receives a plan request that it can partially meet,
controller that is intended to cater for this kind of intended op- the secondary plan request passed to its upstream agent is modi-
eration arc now discussed. fied to reflect this. Within this planning framework, the operator
The agents implemented for the refrigeration system are is the agent furthest downstream and the surface agent is the
shown in Fig. 2. Three supervisory-level agents are defined, one agent furthest upstream. If a plan cannot be fully met, e.g., be-
each for the physical divisions: surface, underground, and de- cause of insufficient supply, a “next best” plan is proposed. This
mand (outlined in Fig. l). Note that all the agents in Fig. 2 do not plan could be rejected by the operator.
always exist in the distn buted system. They are added and re-
moved dynamically, as necessary. Additional Agents
The planning or supervisory agents have subsidiary agents,
Planning Agents which they use to assist in plan formulation or to fulfill the plan.
The three supervisory agents are responsible for system plan- The demand agent is the simplest. It processes operator in-
ning. Each agent implements a rule set that enables planning. puts, i.e., flow and temperature requirements, and starts the de-
Two typical rules coded in the underground agent template are mand pump and control valve when the demand is active.
shown in Fig. 3. Because of its simplicity, the demand agent is actually incorpo-
Because of the distributed nature of the agents and the use of rated into the underground agent. The underground agent has
rules for planning, the planning mechanism resembles a distrib-
uted expert system. Once a demand has been specified by an op- ‘For the sake of planning, the surface dam is considered part of
erator (time interval, flow, and temperature), the superviqory the underground system and agent. This is because the underground
agents cooperate to construct a plan for meeting the demand. agent can satisfy a plan locally if sufficientheated water is available
This plan is influenced by dam levels, ambient temperature, in the surface dam.
April 1998 33
two subsidiary agents, one to calculate the optimal heater con- output (configuration). The input domains have five sets and the
figuration to use and another to control the heater bank. output domain four sets. The logic employed is centered around
The heater configuration is calculated by a configuration the two conditions: all heaters for high flow, high temperature
agent that uses fuzzy logic to implement a decision matrix [23]. and one heater for low flow, low temperature. Given a required
The fuzzy logic has two inputs (flow and temperature) and one flow and temperature, the fuzzy logic returns a required heater
configuration, e.g., two heaters in se-
ries.
The f i r s t i n t h e sequence of demand i n t e r v a l s i s announced. The heater agent accepts a configu-
The underground dam i s n o t capable of s a t i s f y i n g t h e demand ration and water outlet temperature for
l o c a l l y due t o i n s u f f i c i e n t volume; t h e demand agent m o d i f i e s the heater bank. Given the configura-
t h e p l a n t o i n d i c a t e t h e r e q u i r e d volume (flow, t e m p e r a t u r e , tion and outlet temperature, the agent
i n t e r v a l ) and p a s s e s t h e modified p l a n ( a p r o p o s a l ) t o t h e calculates setpoint temperatures for all
underground a g e n t . required heaters and launches PID
The underground agent f o r m u l a t e s a p l a n b u t d i s c o v e r s t h a t t h e agents to control them. The heater
agent includes basic rules. For exam-
h e a t e r i n l e t t e m p e r a t u r e i s t o o low. The underground agent does
ple, if only one heater is required, but
two t h i n g s .
the furthest downstream heater is not
F i r s t it p a s s e s a p l a n p r o p o s a l on t o t h e s u r f a c e agent
available, then the setpoint for the up-
r e q u e s t i n g t h e h e a t e r i n l e t t e m p e r a t u r e t o be r a i s e d .
stream heater is increased to allow for
Second, it r e t u r n s a p l a n p r o p o s a l t o t h e demand agent i n d i c a t i n g
ambient losses in transporting the water
t h a t it can supply t h e demand r e q u i r e m e n t , b u t o n l y a t a reduced
through the downstream heater. The
temperature. magnitude of increase is CO
I n t h e mean t i m e , t h e s u r f a c e agent works on t h e p l a n p r o p o s a l heater agent as approxim
r e c e i v e d from t h e underground agent and f i n d s t h a t it can s a t i s f y grees C.
t h e proposed p l a n r e q u i r e m e n t s . The s u r f a c e agent i n d i c a t e s t o The PID agents implement a stan-
t h e underground agent t h a t it can s a t i s f y t h e proposed p l a n t h a t dard backward difference equation to
it r e c e i v e d . perform PID control [26].The PID con-
The underground agent now i n d i c a t e s t o t h e demand agent t h a t stants and error signal and control sig-
t h e r e q u i r e d demand can be m e t f u l l y . Planning f o r t h e demand nal sources are stored by the heater
i n t e r v a l i s now complete. agent and communicated to newly
launched PID agents. Step tests on each
Fig.4 Planning example heater were used to determine suitable
PID constants.
Experiment 1: The refrigeration plant thus requires
Requirements six agent templates to be defined: de-
mand, underground, surface, heater,
Flow = 2 l/min.
config, and PID. Each agent template
Temperature = 4 5 O C . implements the functionality described
above as well as interactioii rules that,
e Only h e a t e r 1000 a v a i l a b l e ( c o n s t r a i n t ) by capturing the role and behavior of
agents, define when “neighboring”
Results agents are required
Planned h e a t e r c o n f i g u r a t i o n = 1000 with s u r f a c e “ h e a t e r ” p l a n t Consider the underground agent. If
o u t l e t t e m p e r a t u r e set t o 31.8OC and underground o u t l e t it cannot satisfy the plan locally, it will
t e m p e r a t u r e set t o 41.8’C. launch a surface agent. The under-
P l a n Comments ground agent must also launch a con-
figuration agent, and if a valid plan is
I n l e t Temp low -- a t t e m p t i n g t o h e a t . available it will launch a heater agent.
I n l e t being heated. The heater agent only launches the re-
quired PID agents when the under-
Only s i n g l e series h e a t e r a v a i l a b l e -- l i m i t e d maximum Temp. ground agent’s plan becomes active. It
terminates the agents when the plan is
Plan f i n a l i s e d . completed. Similarly, the surface agent
launches and terminates agents as re-
P r i o r t o h e a t i n g i n l e t water temp and a d j u s t i n g for a v a i l a b i l i t y
qurred. The result is that the number of
c o n s t r a i n t s , an underground o u t l e t temperature of 39.2OC w a s
agents in the distributed system
proposed by t h e underground a g e n t . This would hav
changes dynamically to suit operating
t o around 3 5 O C i f t h e s u r f a c e ” h e a t e r ” p l a n t was unable t o h e a t
requirements and conditions
water.
The distributed controller described
Fig 5 Refngerution expenment. Plan 1 vesults above was implemented on a two-node
April 1998 35
launched to heat the inlet water. In developing the final plan, the figuration (i.e., the number o€ PID agents) changed to suit each
comments of Fig. 7 were generated. The final plan called for a demand interval
surface agent to heat the water to 33 degrees C and the two paral- Fig. 8 shows , at a physical level,
le1 heaters 0101 to raise the temperature to 43 degrees C. The flow requrrements were satisfied by the dis
heater outlet temperature setpoints were adjusted to 45.23 de- . For real refrigeration systems, manu
grees C to account for ambient losses. Note that planning and ator is necessary in order to achieve similar results. Note that
plan execution (starting pumps, heaters, operung valves, etc.) heater 0100 is only used to satisfy the second demand interval
were completely automated by the distributed system-only the (higher flow and temperature setpomts). The temperature pro-
intervals and required temperatures were input by the operator files indicate that the initial demand interval temperature set-
The configuration, heater, s ce, and underground agents were point was met approximately 100 seconds after the interval start;
launched on one processi de, while the PID agents were the second interval temperature setpoint was met approximately
launched on both processing nodes. Note that the PID agent con- 150 seconds after the interval start. A small flow disturbance was
introduced into the 0001 heater 900 seconds into the experiment.
The result is a drop (subseqiently re-
covered) in water temperature and a
Heater 0100 Outlet Water Temperature rise in heater power. The water tem-
perature drop soon after the flow distur-
bance marks the end of the first demand
50 ~
20 s due to conduc-
0 500 1000 1500 2000 2500
Time (sec)
heater 0001. Fig. 9 also indicates how
Heater 0100 Control Signal
100 , the surface agent heats the water tem-
I
I
perature in older to meet the second de-
i mand interval. Note that conduction is
I
possible during the time between the
demand intervals because no water
flows during this period
20 Conclusions
Y
tion paradigms
planning and plan mod
suit operating con
0 500 1000 1500 2000 2500 the use of partial results in plan-
Time (sec)
ning, and
100
80
s
v
2 60
6
Q 40
20 (configuration agen
0
0 500 1000 1500 2000 2500
Time iseci
1000
I
1500
I
2000
I
cates the power of a distributedsystem to fully acknowledged. The authors would also like to thank the
facilitate both simple and icomplex pattems of communication.In reviewers for their insightful and constructive comments.
future developments, a more generalized method of transmitting
complete objects [27] between agents could be implemented.
The application of dynamic structuring in the prototype con-
References
[I] B. Chandrasekaran, “Natural and Social System Metaphors for Distrib-
troller proved a successful and useful characteristic, particularly uted Problem Solving: Introduction to the Issue,”IEEE Transactions on Sys-
with respect to the abilitli of the overall system to adapt to the tems, Man and Cybernetics, 11(1):1-5, January 1981.
availability of underground heaters. The authors contend that
[2] W. Kornfeld and C. Hewitt, “The Scientific Community Metaphor,”IEEE
similar levels of flexibilily would be difficult to achieve or im-
Transactions onSystems,ManandCybernetics, 11(1):24-33, January 1981.
practical in control systems based on ad hoc combinations of
conventional controllers. The dynamic structuring comes from [3] C. Harris, guest editorial, International Journal of Control, 56(2):259-
the ability to model the refrigeration system as a cooperating 261, August 1992.
multi-agent system and results in a planning mechanism that [4] M. Rao, “Frontiers and Challenges of Intelligent Process Control,” Engi-
launches specialized agents to suit operating conditions and op- neering Applications of Artificial Intelligence, 5(6):475-481, November
erator demand. The main effect of dynamic structuring in the re- 1992.
frigeration example is to idlocate resources as needed. In a more [5] S. Gershwin, “Opportunities for Control in Manufacturing,” IEEE Trans-
complex controller, where the possibility of mutually exclusive actions on Automatic Control,30(9):833, September 1985.
agents exists, this can be used to good effect. For example, a con-
[6] E. Werner and Y. Demazeau, eds., Decentralized A.I. 3: Proceedings of
troller and fault diagnosis agent would not necessarily need to the Third European Workshop on Modelling Autonomous Agents in a Multi-
exist at the same time. The same computational resource can thus Agent World, North-Holland, 1992.
be used for both functions. Dynamic structuring has an addi-
[7] E. Durfee, “The Distributed Artificial Intelligence Melting Pot,” Transac-
tional important characteristic. It allows the planning mecha-
tionsonSystems,Man undcybernetics, 21(6):1301-1306, November 1991.
nisms to provide partial clr “best-guess’’solutions. For example,
the underground agent might require the cooperation of a surface [8] S. Bird, “Toward a Taxonomy of Multi-Agent Systems,” International
agent to satisfy a plan. The underground agent provides an inter- Journal of Man Machine Studies, 39(4):689-704, October 1993.
mediate solution while the surface agent is launched, and waits [9] L. Gasser and M. Huhns, eds., Distributed Artificial Intelligence Volume
until the surface agent is able to retum a solution of its own. The II, Morgan Kanfmann, 1989.
key point is that workable sub-optimal solutions exist until a final
[ 101G. O’Hare and N. Jennings, Foundations ofDistributedArtifcialIntelli-
solution is agreed to. gence, John Wiley and Sons, 1996.
In summary, the design and testing of a prototype distributed
intelligent controller for a laboratory-scale mine refrigeration [ 111 C. Boutilier, Y. Shoham, and M. Wellman, “Economic Principles of
Multi-Agent Systems,” Artificial Intelligence, 94( 1-2):1-6, July 1997.
system has indicated that the multi-agent approach to control has
several interesting characteristics and is worthy of further devel- [12] E. Werner, “Cooperating Agents: A Unified Theory of Communication
opment for practical, su pervisory-level control of large-scale, and Social Structure,” in L. Gasser and M. Huhns, eds., Distributed Artificial
complex systems. The technique also has the potential to support Intelligence Volume II, chapter 1, pages 3-36. Morgan Kaufmann, 1989.
implementation of advaxed features such as conflict resolution, [ 131 R. Englemore and T. Morgan, Blackboard Systems, Addison-Wesley,
learning, and diagnostic agents. 1988.
[ 141 B. Hayes-Roth, “A Blackboard Architecture for Control,” Artipcial In-
Acknowledgements telligence, 26(3):251-321, July 1985.
The support of the South African Foundation for Research [I51 R. Davis and R.G. Smith, “Negotiation as a Metaphor for Distributed
Development and the University of the Witwatersrand is grate- Problem Solving,” Arrifirial Infelligenre,20:63-109, January 1983.
April 1998 37
[ 161 A. Stothert and I. MacLeon, “DistributedIntelligentControl System for 1271T. Mowbray and R Zahavi, Til? Essential CORRA: ,’$stein.(. Inregrution
a Continuous-State Plant,” IEEE Transactions oil Systems Mail and Cyber- Using Distributed Objects, John Wiley B Sons, 1995.
netics. 27(3):395-401,June 1997.
[ 171 A. Stothert. Coordinating Agent Interaction in Distributed Intelligent Ian MacLeod received the B.Sc.(Eng) (cum laude) in
Coi?trolSystems. Ph.D. thesis, University of the Witwatersrand, Johannes- electrical engineering and the Ph.D.. from the Univer-
burg, March 1996. sity of the Witwatersrand in 1973 and 1983. Afier sev-
eral yea-s in the British and South African chemical
[18] M. Torngren and J. Wilkander, “A Decentralization Methodology for
industries as a control and instrumentation engineer, he
Real-Time Control Applications.” Control Engineering Practice. 4(2):219-
joined the University of the Witwatersrand in 1981.
228,1996.
Since 1984 he has been Fuchs Professor of Control En-
[19] A. Stothert and I. MacLeod, “Research Issues in the Dynamic Structur- gineering in the Department of Electrical Engineering
ing of Distributed Intelligent Controllers,” in IFAC Youth Automation Con- at the University of the Witwatersrand, Johannesburg,
,ference,pp. 605-612, Beijing (China), August 1995. IFAC. Soulh Africa, where he teaches courses on control engineering. industrial
process control, and real-time computer control systems and direcis research
[20]A. Stothert and I. MacLeod, “Using Intelligent Agent Templates for Dy- in the fields of distributed computer control systems and engineering applica-
namic Structuring of Distributed Computer Control Systems,” Engineering tions of artificial intelligence. He is a member of the South African Institute
App/icoliou ofAi.t$cial Intelligeizce, 10(4):335-343,August 1997. of Electrical Engineers, the South African Institute of Measurement and
[21] J. van der Walt and A. Whillier, “Considerations in the Design of Inte- Control, and the Computer Society of South Africa. He is aregistered profes-
grated Systems for Distributing Refrigeration in Deep Mines,” Jo~ii.?ialof the sioud engineer. He has served on the program committees of many interna-
Mine Venrilation Society ofSouth Africa, 31(12):217-243, 1978. tional conferences and is an associate editor of the Transaction of the South
AfricanInstitute ofEIectrica1 Engineers and an advisory editor for Engineer-
[22]M. Bailey-Mchwan, “Use of the Chilly. C‘ .. .. >; , . I. . . . i:Con- ing Applications ofArt$cial Intelligence and Coutrol Engineeriizg Practice.
ventioiial Water Chilling Installations on SOII,; , . , i. 1 I. /. ‘iI . . Jour-
”
Jan J. Kok received the M.E. and doctor of engineer- I van den Steen studied electrical engineering at
I.4)n
ing degrees in mechanical engineering from the Delft ’ :University of Technology Delft and received his
University of Technology. the Netherlands, in 1968 . ” .D. at the University of Twente, both in the Nether-
and 1978. respectively. From 1965 until 1985 he I .inis.After working in various areas of research at
worked in the Man-Machine Systems Group of the De- I’ #lipsResearch and TNO Medical Biological Labora-
partment of Mechanical Engineering of the same uni- .. f, he joined Shell Research, currently calledshell In-
versity. Since 1985 he has been professor in systems tational Exploration and Production, in 1981 to
and control in the Department of Mechanical Engi- ’ form research on subsea control systems for oil and
neering of the Eindhoven University of Technology, 3 production. Since1990 he has been involvedindrill-
and since 1991 also in the Department of Technical Physics. His interests string dynamics research and, in particular. in the active damping of torsional
include (nonlinear) modeling, identification and control of electrome- drillstring oscillations.
chanical systems and its applications.