You are on page 1of 6

A Multi-Agent Approach to Unit Commitment

Problems
T. Nagata, Member, IEEE, M. Ohono, non-Member, J. Kubokawa, Member, IEEE,
H. Sasaki, Member, IEEE and H. Fujita, Member, IEEE

optimization problems. This problem is one of mixed integer


Abstract This paper proposes a multi-agent system for programming which consists of 0-1 variables that stand for
solving unit commitment problems. Multi-agent is a new on and off status as well as the real power output. The
paradigm for developing software applications. Coordinating proposed methods so far are classified into the heuristic
the behavior of autonomous agents is a key issue in agent method, the dynamic programming [1], the Lagrange
oriented programming techniques today. In this paper, we relaxation method [2, 3], the branch-and-bound method [4,5]
develop a power system unit commitment application using and the evolutionary programming method [6-9]. We have
multi-agent architecture. Our model has the following studied UC solutions using genetic algorithms and the
characteristics: (1) the system consists of a single facilitator
Lagrange relaxation method in consideration of LNG fuel
agent, several generator-agents, and two kinds of mobile agent.
(2) The facilitator agent is developed to act as a manager for the
constraints [10] with successful results. Since UC in the new
process by using the singleton design pattern. The mobile agent deregulated business environment requires the coordination
and generator agent have simple negotiation strategies. (3) of all market participants including IPPs, it is difficult to
Message object is developed to communicate between agents collect all information from these parties because of the
using KQML-like object. The proposed approach is applied to a commercial secrets. In this situation, uncertainties for
simple system, and the results show that the multi-agent system utilization of the generators will be increased, since the
is able to find optimal solutions for unit commitment problems. inspection and maintenance work of IPPs generators will be
carried out on the basis of owners decision making. These
Index TermsAgent technology, Mixed Integer Program- considerations indicate that UC must be decided by obtaining
ming, Multi-Agents, Unit Commitment Problem the latest (on-demand) information from IPPs. In other words,
it means that UC must be decided on the assumption that all
I. INTRODUCTION the information comes from the distributed server computer
at each IPPs site.
T HE demand of intelligent information processing is
heightening more and more in proportion to changes in
the business environment surrounding electric power systems
Most research of UC problem were based on the
concentrated (centralized) systems, where there is little
such as energy problems, environmental problems and research based on distributed systems.
deregulation, and in proportion to the need to handle In this paper, we will propose a UC method using a multi-
complicated, enlarged, decentralized and opened power agent approach on the distributed system environment.
systems. With the promotion of deregulation of electric Although we assume that the proposed system would be
power systems and participation of independent power implemented on the distributed servers, we have placed
producers (IPPs), it is becoming extremely difficult for the several agents onto single computer system for simplicity in
power system operator to handle all control information into the prototype system. The prototype multi-agent system
a single control center computer. Therefore, paradigm change solves minimization problem of total operational cost by
in control system is required from a concentrated to a traveling a mobile agent to each generator agent.
distributed system, in line with what has happened in other
fields in present-day society and industry. The features of the proposed method are:
Unit commitment (UC) is the problem of determining the (1) In order to solve the problem efficiently, we restrict the
start-up and shut-down schedule of thermal units to meet kind of agent which carries out the decision making
forecasted demand over certain time periods ahead (24 hours process, and furthermore reduce the number of message
to a week) and belongs to a class of combinational exchanges as few as possible during UC scheduling. The
proposed system consists of three types of agent: a single
facilitator agent (FAG), several generator agents
Takeshi Nagata, Masumi Ohno and Junji Kubokawa are with Hiroshima (GAG) and two mobile agents.
Institute of Technology, 2-1-1 Miyake, Saeki-ku, Hiroshima, 731-5193 Japan, (2) Two types of mobile agents have been introduced to
(e-mail: nagata@cc.it-hiroshima.ac.jp, kubokawa@me.it-hiroshima.ac.jp)
Hiroshi Sasaki is Professor of Hiroshima University, 1-4-1 Kagamiyama,
improve the global search efficiently. That is, up
Higashi-Hiroshima, 739-8527, Japan (e-mail: sasaki@psl.sys.hiroshima-u.ac.jp) mobile agent (UpAG) with the function of increasing
Hideki Fujita is with Research & Development Center, Chubu Electric Power power generation output, and the down mobile agent
Co. Inc, 20-1,Kitasekiyama Ohdaka-cho Midori-ku, Nagoya, 459-8522, Japan
(e-mail: Fujita.Hideki@chuden.co.jp)
(DwAG) with the function of decreasing power
0-7803-7322-7/02/$17.00 2002 IEEE
64
0-7803-7322-7/02/$17.00 (c) 2002 IEEE
generation output. In the proposed multi-agent system,
B. Formulation
depending on the demand-generation balance condition,
either UpAG or DwAG travels around the network and a) Objective Function
negotiates with GAGs. UC problems can be formulated as an optimization
(3) GAGs, which would be placed on the server computer at problem with an objective of minimizing the total operational
each generator/IPPs, hold a few simple rules and update cost subject to some constraints.
T I
the UC status (on/off) and the power output min C(pit ,u ti ) = min c i( pti ,u ti )
t t
autonomously. t
u i, pi
t
u i, pi t =1 i =1

(4) FAG notifies the status of all GAGs to the black board,
( )
T I
t 1
= min FC i pi ui + ST i ui 1 ui
t t t
sends the request message of the power regulation to
u , p t =1 i =1
t t
i i
DwAG or UpAG, and waits for the report from the
mobile agent concerned. The singleton design pattern (1)
t
[11] is adopted in the proposed system, because only one where C( pi , uti) corresponds to the total operational cost and
FAG can be exist on a center server connecting to other can be divided into each unit.
distributed servers. T I

(5) The message exchange between agents is performed by C( pti , uit ) = c i(pit ,u ti ) (2)
t =1 i =1
using a message object that is similar to KQML
In General, the mathematical formulation of UC problem
(Knowledge Query and Manipulation Language).
is to minimize two kinds of cost. The first term in (1)
indicates the production cost, which is directly related to the
The proposed multi-agent system has been implemented on a
fuel cost and is formulated as:
PC with Java language, and applied to a 10-unit model t
system. The result shows that the proposed system is able to FC i (pit ) = a i0 + ai1 pi + ai2 (pti )2 (3)
obtain solutions almost the same as the dynamic where aij is the coefficient of the fuel cost.
programming method [12], which minimizes the total The second term corresponds to the start-up cost for each
operational cost. The solution time is about 70 seconds unit. The cost of this term is related to the temperature of the
including the display of intermediate results (PC, 550MHz). boiler. The cooler the boiler temperature leads to the higher
the start-up cost. Therefore, the start-up cost of each unit can
II. FORMULATION OF THE UNIT COMMITMENT PROBLEM be simply defined as the following formulation.
HSC i if MDT i < DT i MDT i + CSH i
A. Explanation of Variables ST i = (4)
CSC i if DT i > MDT i + CSH i
To formulate the problem mathematically, the following
notation is introduced:
i : Index of generator ( i = 1,L, I ) b) Constraints
t : Index of period ( t = 1,L,T ) There are two types of constraints that must be satisfied by
l : Index of transmission line ( l = 1,L, L ) a UC schedule. Those affect units as a group is called
ui
t
: Commitment status (1 or 0) of unit i at period t system or coupling constraints while those constrain the
t operation of each unit individually are termed local or
pi : Power generation of unit i at hour t
horizontal constraints. Here, in the following constraints,
ci : Power system operating cost of unit i (5) and (6) are the system level, and (7) and (8) are the local
FCi : Fuel cost of unit i ($) level constraints.
ST i : Start-up cost of unit i ($) Supply and Demand Balance Constraint
t
D : Forecasted demand at period t (MW) This constraint ensures that the total power generation
R
t
: Spinning reserve at period t (MW) output is equal to the forecasted demand at each period.
I

p
max t
pi : Upper power limit of unit i (MW) i
D t = 0 t = 1,L, T (5)
min i=1
pi : Lower power limit of unit i (MW) Spinning Reserve Constraint
HSCi : Hot start cost of unit i ($) The spinning reserve constraint is necessary to maintain
CSC i : Cold start cost of unit i ($) the reliability of the power system. Normally, the spinning
MDT i : Minimum down time of unit i (h) reserve requirement is given by the ratio (%) of the total
MUT i : Minimum up time of unit i (h) demand at each period.
I

u
max
DT i : Time for which unit i has been off-line (h) t
pi Dt R t 0 t = 1,L, T (6)
i
CSH i : Cold start time of unit i (h) i=1

65
0-7803-7322-7/02/$17.00 (c) 2002 IEEE
Generation Lower and Upper Limit Constraints
The power generation output of each generation must be migrate
GAG
ranged between the lower and upper limits of generation
capacity. request/inform migrate
DwAG
min t max
u i pi pi u i pi
t t
(7) FAG GAG
Technical Constraints UpAG migrate
request/inform
The technical constraints of minimum start-up time and migrate
minimum shut-down time are considered in this research. GAG
Fig. 1 A multi-agent unit commitment system
III. UNIT COMMITMENT MULTI-AGENT SYSTEM
In this section, we explain an approach that achieves UC, A. Explanation of Composed Agent
as described previously, by using a multi-agent system. Fig. 1
shows the structure of the proposed multi-agent system for In the following, we explain about agents that comprise the
UC. This figure shows that the multi-agent system consists of proposed multi-agent system.
three generator agents (GAGs), a single facilitator agent
(FAG) and two types of mobile agent (DwAG and UpAG). (1) Decreasing production mobile agent (DwAG)
First, the objective function minimization can be achieved DwAG is produced and started up by FAG when the
by the mobile agent (UpAG or DwAG), which minimizes the demand shows a tendency to decrease. This agent migrates to
total operational cost. The mobile agent, which is requested GAG along the traveling route made by FAG, decides the
to make a schedule from FAG, goes around the traveling production schedule, and informs FAG. Here, the schedule is
route of the network, and decides the generation schedule by decided by performing a message exchange with GAGs.
using the strategy that each unit can generate power with up (2) Increasing production mobile agent (UpAG)
to its maximum value in the order of increasing the full-load UpAG is produced and started up by FAG when the
average production cost [12], which is adopted frequently in demand shows a tendency to increase. In the same manner as
actual UC with heuristic methods. The full-load average DwAG, this agent migrates to GAG along the traveling route
production cost is expressed as follows: made by FAG, decides the production schedule and informs
max FAG.
a i0 + ai1 pi
FLACi = max (8) (3) Facilitator agent (FAG)
pi FAG produces either DwAG or UpAG in response to the
From the above equation, the full-load average production condition in the decision making process (to satisfy demand-
cost is an approximation because it doesnt include the non- generation balance constraints at the consideration period).
linear fuel cost characteristic of the unit. Note that there is a Moreover, this agent determines the traveling route from the
case that doesnt have the cheapest operational cost at any units operational status registered on the Black Board, and
period, since each unit generates the power with maximum sends the message with the traveling route attached, which
output value in the ascending order of the full-load average starts the scheduling process, to the corresponding mobile
production cost in the proposed method. agent.
On the other hand, the system constraints (i.e., the supply (4) Generator agent (GAG)
and demand balance constraints, and the spinning reserve GAG is the agent corresponding to the generators / IPPs in
constraints) will be met after the mobile agent migrates to the electric power system. This agent has fixed data such as
GAG along the traveling route and negotiates with it. The unit name, the minimum and maximum power, the fuel cost
local constraints (i.e., the lower and upper limits, and the coefficient, and variable data such as the power production
minimum start-up time, the minimum shut-down time) will value and the status of the unit at the current time.
be satisfied by individual GAG. UC scheduling process is B. Explanation of Messages
started by informing scheduling request from FAG to either
DwAG or UpAG, and the concerned mobile agent decides The communication between agents in the system is based
the schedule by migrating to each GAG. Here, DwAG is on a message object similar to KQML, as shown in Fig 2. In
started up when the demand shows a tendency to decrease. this figure, each message consists of performative, contents,
On the contrary, UpAG is started up when the demand shows sender, receiver and arguments. Table 1 shows the major
a tendency to increase. performative and its meanings used in the proposed system.

66
0-7803-7322-7/02/$17.00 (c) 2002 IEEE
PerformativeContentsfrom Sender to GAG1
Receiver withArguments UpAG 150-600
500->600
Fig. 2 Structure of message object
migrate
TABLE 1 PERFORMATIVE AND ITS MEANING UpAG GAG2
UpAG 100-400
Performative Meanings FAG +200 0->200

Request Execution of the request action


GAG3
Query Query for information 50-200
0
Inform Inform on information
Propose Execution of the proposed action (b) Migration from GAG1 to GAG2
Agree Execution of the agreed action GAG1
150-600
Refuse Execution of the refused action UpAG 500->600

IV. IMPLEMENTATIONS inform UpAG GAG2


UpAG 100-400
FAG +200 0->200
In order to illustrate the proposed scheme, we will use a
migrate
simple model system shown in Fig 1. For the sake of
GAG3
simplicity, this simple system considers only the generation 50-200
0
and demand balance constraint. Table 2 and 3 show the data
of GAGs and the forecasted demand respectively. The (c) Migration from GAG2 to FAG
traveling route is in the order of GAG1, 2 and 3 Fig. 3 An explanation of mobile agent
corresponding to the full load average cost.
(Step1) Migration to GAG1
TABLE2 GENERATOR DATA TABLE 3 DEMAND DATA First, for the purpose of increasing generation 300 (MW),
GAG1 GAG2 GAG3 Time (MW) UpAG migrates to GAG1 with the highest priority along the
Pmin (MW) 150 100 50 1 1000 traveling route and sends a message (request), which prompts
Pmax (MW) 600 400 200 2 750 an increase in generation, to GAG1. Then, GAG1, having
a0 ($) 510 310 150 3 500 received the request message, performs power generation
a1 ($/h) 7.2 7.85 7.97 4 800 output adjustment of +100 (MW) in order to increase the
a2 ($/h*h) 0.00142 0.00194 0.00482 power generation, and returns an agree message (agree) to
MUT(h) 5 4 3 UpAG. In this case, UpAG migrates to GAG2 to make up for
MDT(h) 5 1 1 the shortage of power generation (200MW) with the aim of
CSC($) 9000 1100 900 increasing production. Fig. 3(a)-(b) show the above-
HSC($) 4500 550 450 mentioned situations. If GAG1 is working at maximum
CSH(h) 5 4 3 power, GAG1 returns a refuse message (refuse) to UpAG
Here, we explain about the action of the mobile agent in because GAG1 cannot increase power generation. In this case,
the case of an increase in demand. For example, at period 4, UpAG also migrates to the second priority GAG2 for the
it is the case that the demand increases from D(3)=500 (MW) purpose of increasing generation. In both cases, if the power
to D(4)=800 (MW). At this point, because of a tendency for generation meets the demand because of output adjustment of
demand to increase, UpAG is produced and started up by each GAG, UpAG migrates to FAG and reports the result to
FAG. FAG.
In what follows, we outline the action of UpAG using Fig.
3. At period 3, GAG1 is only on status and its power (Step2) Migration to GAG2
generation is 500(MW). UpAG migrates to the second priority GAG2 as shown in
Fig 3(b). After completing the migration, UpAG sends the
request message for start-up (request) to GAG2 because
GAG1
migrate UpAG 150-600 GAG2 is shut-down. GAG2, which receives the message,
+300 500->600 changes the status from off to on, adjusts power generation
request
D(4)=800 output to 200 (MW), and returns an agree message (agree) to
UpAG GAG2 UpAG. In this case, UpAG migrates to FAG and informs a
100-400
FAG 0 successful result, because power balance between generation
and demand has been met.
GAG3
50-200
0 (Step3) Migration to FAG
(a) Migration to GAG1
At period 4 (time=4), since the scheduling process is
completed, UpAG migrates to FAG and informs the results.
FAG, which receives the results, also produces either UpGA

67
0-7803-7322-7/02/$17.00 (c) 2002 IEEE
or DwGA depending on the power balance between
generation and demand at the next period, and starts it up. TABLE 4 SIMULATION RESULTS
(a) Case of ignoring the start-up cost and the technical constraints
In case of decreasing power generation agent DwAG is
GAG1 111111111111111111111111
produced and started up by FAG, the same processes as
GAG2 111111111111111111111111
increasing case are performed except for the traveling route.
When these steps are carried out over the time considered GAG3 000001111111111111111100
under the control of FAG, UC process is completed. GAG4 000111111111111111111100
Meanwhile, we can consider the above-mentioned processes GAG5 000000011111111000111000
as a contract of power trade (buying and selling) by GAG6 000000000111100000011000
negotiating between the mobile agent and GAGs. It is a key- GAG7 000000000011000000000100
point for the proposed system that orientates the distributed GAG8 000000000001000000000000
system to reflect uncertainty, such as the availability of GAG9 000000000000000000000000
generators and the amount of power generations for sale, by GAG10 000000000000000000000000
negotiating between agents.
Cost ($) $545,009 (+0.07%)
DP $544,636
V. SIMULATION

A. Simulation Conditions (b) Case of with the start-up cost and the technical constraints
GAG1 111111111111111111111111
In order to demonstrate the performance of the proposed
GAG2 111111111111111111111111
system, a prototype code has been implemented on PC with
Java language and applied to a 10-unit model system. The GAG3 000001111111111111111000
generator data and the forecasted demand data are shown in GAG4 000111111111111111111100
reference [7]. The traveling route is in the order of GAG 1, 2, GAG5 000000011111111000000000
4, 3, 5, 6, 7, 8, 9, and 10 corresponding to the full load GAG6 000000000111000000111000
average cost. In order to simplify the simulation, the spinning GAG7 000000000011100000011100
reserve constraints are removed. GAG8 000000000001000000010000
B. Simulation Results GAG9 000000000000000000010000
Table 4 shows the simulation results of a 10-unit model GAG10 000000000000000000000000
system over a 24-hour period. In this table, 0 and 1 denote Cost ($) $556,080 (+0.37%)
the off and on status respectively. We have compared the DP $554,041
proposed method with the dynamic programming method
(DP) [12] to confirm the accuracy of solutions. Table 4(b) shows the results in the case of considering both
First, table 4(a) shows the results in the case of ignoring the start-up cost and the technical constraints. The obtained
the start-up cost and the technical constraints (i.e., the result for the proposed method is 0.37% higher than the result
minimum start-up time and shut-down time). The on-off of DP method. There is a possibility to have lower total
status of UC obtained in this case is the same as the results of operational cost for the proposed method, if we apply proper
DP method. However, the total operational cost is 0.07% load dispatching method to the committed generators. The
higher than the result of DP method. This difference is owing half-toned part of the Table 4(b) indicates the difference
to the method of the load dispatching, that is, since the pro- between the proposed and DP method. These differences
posed method tries to allocate the generation up to the comes from the situation that, in the proposed system, the
maximum value of the on-line unit in order of traveling route, mobile agent tries to start up units along the traveling route
optimally economical loading dispatching is not always and migrates to the next GAG when the minimum down time
realized. If the proposed method is modified so that the is not satisfied, while DP method stores all available
mobile agent is able to decide the proper loading (i.e., solves transition paths at the forward process and selects the optimal
the economic load dispatching problem) after commitment path for the backward process. In other words, the proposed
status is fixed and informs the results to each GAG, it may method does not search all available transition paths.
achieve the same results as DP method. Although the proposed method could certainly add such
functions to the mobile agent, it is meaningless for the
problem of the practical size because it is anticipated easily
that the revised method falls into the situation such as the
curse of dimensionality of DP.
From the above results, it can be seen that the proposed
method can obtain solutions comparable to the results of DP
method, although there is room for improvement of the
proposed method. The computational time for a 10-unit over
a 24-hour period is about 70 seconds including the output of

68
0-7803-7322-7/02/$17.00 (c) 2002 IEEE
the intermediate results (PC, 550MHz). We have assumed [12] A.J.Wood, B.F.Wollenberg, Power Generation Operation and Control,
that GAG is placed on the server computer of each power John Willey & Sons, Inc. (1996)

producer (including IPP) in practice, though in this


simulation each GAG is arranged in the same computer for
simplification. Since the proposed method adopts the RMI,
the Class-Loader, and the Object Serialization Library
provided in Java environment, the mobile agent in this
system can migrate to a different server computer.

VI. CONCLUSIONS
In this paper, we have proposed an approach that models the
decision making process of UC with a multi-agent system
and implements it as the first step to examine the possibility
of using a distributed UC system in the new deregulated
electric power system business environment.
We have utilized three types of agent (a single Facilitator
agent, several Generator agents and two kinds of Mobile
agent) in the proposed multi-agent system. We have
confirmed that the obtained results are comparable to the
results of DP method for total operational cost minimization
problem. Therefore, though more detailed examination is
necessary in the future, this paper can be regarded as a
contribution to one way of implementing a distributed UC
multi-agent system.
In the future, we will expand the proposed system into a
new multi-agent system suitable to the new business
environment of the electric power system that not only
minimizes total operational cost but also takes into
consideration maximization of the profit of each power
producer.

[1] W.L. Snyder,Jr, et al., Dynamic Programming Approach to Unit


Commitment, IEEE Trans. on Power Systems, Vol. PWRS-2, No. 2,
339-350 (1987)
[2] A.Merlin, P.Sandrin, A New Method for Unit Commitment at
Electricite de France, IEEE Trans. on Power Apparatus and Systems,
Vol. PAS-102, No.5, 1218-1225 (1983)
[3] X.Guan, et al., Nonlinear Approximation Method in Lagrangian
Relaxation-Based Algorithms for Hydrothermal Scheduling, IEEE
Trans. on Power Apparatus and Systems, vol. 10, No. 2, 772-787 (1994)
[4] G.S. Lauer, et al., Solution of Large-Scale Optimal Unit Commitment
Problems, IEEE Trans. on Power Apparatus and Systems, Vol. PAS-
101, No.1, 79-86 (1982)
[5] A.I. Cohen, M. Yoshimura, A Branch-and-Bound Algorithm for Unit
Commitment, IEEE Trans. on Power Apparatus and Systems, Vol.
PAS-102, No.2, 444-451 (1987)
[6] H.Sasaki, et al., A Solution Method of Unit Commitment by Artificial
Neural Networks, IEEE Trans. on Power Apparatus and Systems, Vol.
7, No. 3, 974-981 (1992)
[7] S.A.Kazarlis, et al., A Genetic Algorithm Solution to the Unit
Commitment Problem, IEEE Trans. on Power Apparatus and Systems,
Vol. 11, No. 1, 83-92 (1996)
[8] T.T.Maifeld, G.B.Sheble, Genetic-based Unit Commitment Algorithm,
IEEE Trans. on Power Systems, vol. 11, No.3, 1359-1370 (1996)
[9] D.Kuokka, L.Harada, On Using KQML for Matchmaking,
Proceedings of First International Conference on Multiagent Systems,
AAAI Press (1995)
[10] T. Nagata, H. Sasaki, H. Fujita, An Efficient Solution Method for the
Unit Commitment Problem Considering LNG Fuel Constraints, Large
Engineering Systems Conference on Power Engineering (LESCOPE-99),
91-95, (1999)
[11] E. Gamma, et al.,Design Patterns: Elements of Reusable Object-
Oriented Software, Adison-Wesley (1995)

69
0-7803-7322-7/02/$17.00 (c) 2002 IEEE

You might also like