Artificial Intelligence in Power System

Operations

Power system operators often reach a cognitive barrier when

information arrives too fast during a power system emergency.At
such times it becomes difficult to reach a correct diagnosis of the
problem or to formulate the correct decision when actions must
be taken. Artificial Intelligence gives designersof Energy Management Systems a way to solve many of the diagnosis and decision
problems so as to make the EMS more useful. This paper explores
reasons why AI techniques, such as knowledgebased expert systems, arebeing used in EMS designs and the differences between
knowledgebased expert systems and traditional numeric a l p
rithm development. The differences between expert systems and
the numeric approach extend to the basic conception and design
of the applications. This is illustrated using a relay fault diagnosis
system, showing boththe
traditional and rapid prototyping
approaches to its development. Finally, issues concerned with the
implementation of AI in EMS computers are explored along with
the authors' predictions of possible AI applications to power system operations.

I. INTRODUCTION

Modern power systems are operated by highly skilled

operators through computerized control systems. The
energy managementsystem (EMS)is the center of a control
system organized in a hierarchical structure utilizing remote
terminal units, communication links, and various levelsof
computer processing systems. The function of the EMS is
to ensure the secure and economic operation of thepower
system as well as to facilitate the minute-by-minute tasks
carried out by the
operations personnel. The EMS is mainly
designed to be used in the"norma1"state where such functions as state estimation, security analysis, and optimal
power flow are usedto ensure secureoperation while functions such as automatic generation control, economic dispatch, unit commitment, and load forecasting are used to
ensure that the most economic operation is obtained. Much
of what happensin normal operation is now computerized
and human operators only intervene to carry out the few
manual tasks required.
Manuscript received November 17,1986; revised May 10,1987.
6. F. Wollenberg iswith Control Datacorporation, Minneapolis,
MN 55441, USA.
T. Sakaguchi i s with the
Central Research Laboratory,Mitsubishi
Electric Corporation, 1-1, Tsukaguchi Honrnachi 8 Chome, Arnagasaki, Hyoga 661, Japan.
I E E E Log Number 8717877.

The picture is quite different, however, when we look at

the use of an EMS during an unforeseen event or a failure
of major components on the power system. In such
instances, the EMS serves mainly as an information gathering and reporting system and the sophisticated apptication s~ftware
that functions in normal operation may be
of little use. For example, in a regulatory shutdown of all
nuclear units, human operators will take overtheeconomic
dispatch and unit commitment to resc:hedule economic
operation. Similarly, when a sudden loss o f transmission
equipment occurs it is human operators who must understand what has happened and decide on what actions to
take. It is especially during such emergenciesthat conventionalsoftwareis1esseffective.Therequirementforsmarter
software thus becomes more important in such instances.
Coping with emergency events is reflerred to as a diagnosis and decision process.Suchprocessesare
illstructured and their solutionrests heavily on the experience and
skill of the human operators to react correctly. The key to
human ability in such situations lies in the experience with
sirnilareventsandtheuseof heuristicstomapexistingcomplex situations onto learned past eventsto solve problems.
Since Artificial Intelligence allows the realization of heuristic techniques in a computer, the waly is now open for
many new applications of computers in power system operations.
There is also a need to incorporate heuristics in many of
the functions now carried out by large application programs. For example, security analysis programs may fail to
converge or economic analysis programs may fail to meet
certain operating constraints because the constraints are
difficult to express mathematically. In such cases, human
intervention must be relied on tosolve or circumvent the
problem. Research usingartificial intelligenceisalso beginning to yield methods of embedding complex heuristics
into conventional application software for use in normal
power system operations.
In this paper we are specifically addressing the use of
expert or knowledge-based systems wlhich is one of the
principal branches of artificial intelligence. Several experimental and near-practical-use expert systems have been
developed for use in power system operations. Sakaguchi

oola~191~7/120(r16mw1.m
o 1987 IEEE

1678

PROCEEDINGS OF T H E IEEE, VOL. 75, NO. 12, DECEMBER 1987

and Matsumoto [Idiscuss

]
the use of an expert system to
capture heuristic knowledge from operators and operations manualsthat direct thesteps taken in restoring power
after a large system failure. Wollenberg [2] and Larsoneta/.
[3] describe an experimental expert system to intercept
alarm messages and present the operator with a summary
of the most pertinent information.The authors in [4]-[8] discuss the use of expert systems in application programs.
Power systemoperations is not tlhe only real-time control
environment where expert systems are being applied. References [9]-[Ill describe the use alf expert systems in process control applications and [I21describes an experimental expert system to aid nuclear plant operators in
diagnosing reactor shutdown prolblems. Several common
features can be found inthese systems. First,they normally
require human operators 24 h aday and this gives a great
advantage to an expert system that can provide the right
expertise at any hour and any day of the week. Secondly,
the expert system can potentially provide a rapidreaction
to emergency events by summarizing information quickly
and checking many more applicable rules than a human
operator could in thesame period of time.
This paper begins by analyzing the diagnosis and decision process in power system operiltions and discusses how
this process might benefit from thle use of expert systems.
After describing typical examples, efforts are madeto define
what the AI techniques are and how they differ from
numeric techniques.such as linear andnonlinear programming.The strength of AI overconventional techniques and
how knowledge-intensive problerns are solved is also discussed. Early experiences so far have indicated that implementation of AI in EMS computers poses a difficult problem.Several alternate implementation schemesare proposed andexplored. Finally, some predictions about future
uses of AI in E MS are madeand sorne needed technological
innovationsare listed that are required before AI candeeply
penetrate power system operations.

11. DIAGNOSIS
AND DECISION
PROC:ESSES
IN POWERSYSTEM
OPERATIONS

Is there a need for knowledge-based software in power

systems operations? We believe there is. A fundamental
motivation for such software is the need to overcome the
human cognitive barrier which conventional E M S installations encounter during emergency operation or when
application programs are used beyond their design limitations,
The cognitive barrier is felt as t:he complexity of power
system operations increases without sufficient efforts to
cope with it. This is true of todaysE M S installations where
the quantityof data gathered and the rate at which they are
gathered can overwhelm a humcanoperator. It must be
noted that the driving force in EMS complexity is the desire
to operate the power system closerto its limits so as to make
better use of generation and trarwnission facilities. This,
in turn, has madea qualitative change in system operations
requiring quicker diagnosis and decision making by operators. Fig. 1 illustrates this situation. While system complexityincreasessteadily,theoperatorsabiIitytocopewith
it decreases. Since the complexity of power system operations is very likely to continue to increase in the future,

WOLLENBERG ANDSAKACUCHI: AI IN POWERSYSTEM

OPERATIONS

IS SYSTEM COMPLEXITY
/

Lb

1
1

T Y

Fig. 1. The risk of enlarging the human cognitive barrier.

there is a risk of human operators being unable to manage

certain functions unless their capability is enhanced.
As indicated earlier, the cognitive barrier is quickly realized in power system operations when sudden and unforeseen events occur. When human operators meet such an
event, they have to understand the situation (diagnosis)and
determine actions (decision) to return the system to normal. As all the tasks haveto be done in real time,the operators are exposed
to heavy mental stress and this makes the
cognitive processes distinctively different from that experienced by others. For example, experienced designers of
large-scale integrated circuits often face similar cognitive
barriers but solutions are not required in real time.
There are several waysto help operators overcome this
cognitive barrier. First, operators need to understand what
is happening on the power system and the AI software can
give guidance by showing various scenariosthat explain the
situation consistently. Operators can then check for the
most plausible scenarios, some of which may have been
overlooked. Similarguidancecan beexpected in using large
application programs where the AI system can guide the
operator in its use. Further, sincepower system operations
are filled withmany fragmented tasks that are done almost
routinely, smart software to do such routine tasks could
relieve operators and allow them to devote their time to
more important tasks.
Ill. KNOWLEDGEINTENSIVE
VERSUS NUMERIC
PROBLEMS

An expert system is a software paradigm where knowledgeconcerningacomplexproblem isencoded intoacomputer program. The framework of expert systems is
designed to enable easy encoding of knowledge and easy
checkout of the expert systems performance. A general
architecture for expert systems is shown in Fig. 2. Four major
software elements comprise an expert system: the knowledge base, aninference engine, building and checkout utilities, and the user interface.
In order to use a knowledge base on a computer some
facilities are needed to read a module from the knowledge
base, decide whether it is to be executed,and to carry out
the execution. The inference engine is responsible for this
task and functions much like an interpreter for conven-

1679

called a production rule, which has a premisepart (IFclause)

and an action-part (THENclause) as shown below:
(p

rule-name
(premise-part)

(action-part) )

The rule-basedsys!:em hastwo kirldsof memory: short-term

(or working memory) and long-term. The short-term memory(STM)containsfactual knowledge, to be modified as the
computation proceeds. The long-term memory (LTM) containsthe productionrules themselves.The inferenceengine
of the rule-based system tests
the pre'mise-partby matching
itagainstthefactualknowledgeintheSTM(matchingcyc1e).
If it succeeds, the action-part of the rule is executed resulting in some changes to theSTM (firing cycle). Theengine
then goes back to the matching cycle. Theremay be more
than one rule which succeeds in matching and the inference engine then invokes a conflict resolution mechanism
to decide which rule shall1 be used.

USER NTERFACE

Fig. 2. Architecture of an expert system.

tional programs. The actual execution or computation in

the inference engine, however, is quite different from an
interpreter. The other software elements are for the convenience of the developer and user of the software. Typical
expert system programs are verylarge and the complexity
of the knowledge base necessitates advancededitors and
browsing tools to debug them. Expert systems alsoprovide
the abilityto explain the reasoning used (e.&, to trace the
rules used in a rule-based system) which is important in
checking it out.
Depending on the representation scheme, anAI program
becomes either rule-based, frame-based, or logic-based as
shown below.
A. Rule-Based System

The popular OPS5 expert system, [13],will be used as an

example. A piece of knowledge is represented in a form

B. Frame-Based System
In the rule-based system, factual knowledge is stored in
the STM without regard to relationships between different
objects. Thereare, however, relations between the objects
of many problems and a frame-based knowledge representation allows the user to set up and make use of these
relationships.
For example, consider the objects of a substation such
as breakers, switches, buses, transformers, and transmission lines. Several objects comprise a substation, and a set
of substations becomes an area. Depending on the status
of individual breakers and switches, buses may be split or
de-energized. Transformers and lines may be connected,
open-ended, or de-energized depending on the status of
the terminatingbus sections, etc. This is illustrated by the
following set of frames:

SUBSTATION FRAME:
character string
Name:
(pointer to breaker frame)
Breakers:
Bus sections: (pointer to bus section frame)
(pointer to lineframe)
Lines:
Transformers: (pointer to lineframe)
BREAKER FRAME:
Name:
Duty rating:
Terminal bus sections:
Status:
LINE FRAME:
Name:
Rating:
Terminal bus sections:
Status:

character string
real constant
(pointers to bus section objects)
openlclosedlunknown
determination: access-breakerdata
character string
real constant
(pointers to bus section objects)
connectedlopenendedlde-energized
determination: line-status-algorithm

BUS SECTION OBJECTS:

character string
Name:
real constant
Voltage class:
energizedlde-energizedlunknown
Status:
determination: bus-section-algorithm
1680

PROCEEDINGS OF THE IEEE, VOC. 75,

NO. 12, DECEMBER

1987

This type of structure allows an expert system that knows

something (i.e., some object has been set to a value in the
STM) about some part of thesubstation (say, the status of
a breaker) to use the frame relationships to directly infer
conditions of other objects in the substation (say, the status
of a line). The Automatic ReasoningTool (ART) expert system language [I41is an example of this type of structure.
C. Logic-Based System
The frameworks we havedealt with so far areappropriate
to represent procedural knowledge such as: what to do
when certain conditions are met. A different way t o r e p
resent knowledge requires one to specify what instead
of how. A logic-basedsystem provides us with such
means. Prolog i s a programming language to represent a
what-type knowledge based on predicate calculus [15].
For example, the following Prolog statement might be
used in an expert system to guide operators in system restoration:
restoration-required( X, yes )
component( X ),
charge-state( X, Y ), =I=( Y, is-charged),
=/=(Z, is-faulted).
fault-state( X, Z ),
+

This example stateswhat must be true for object X in order

for it tohave a restoration requirement of yes.The statement says that X must be a component and that X is not
charged and is not faulted.
Logic-basedsystems have an advantagewhen specifying
system requirements, but they have adisadvantagein specifying procedure-oriented knowledge.
D. How Expert Systems Differ from Numeric Methods

A goodillustrat;on of the difference between expert systems andnumerical methods occurs when one tries to apply
both techniques to the power system restoration problem.
In the case of expert systems, a set of rules which govern
the procedures (or specify conditions to be met in the case
of logic-basedsystems) is written. The inferenceengine then
uses the rules to find a sequence of switchirlg actions to
restore operation to the system. The sameproblem can be
posed asaninteger programming problem and solvedusing
a general-purpose integer programming routine. Two
important points must then be considered, the generality
and the framework of the solution.
Generality has advantages and disadvantages. The fact
that a technique such as integer programming canbe
applied to this problem may not be of much benefit since
the feasible solution space is very large and the restoration
problem does not meet the convexity assumptions needed
to guide the selection of alternative solutions. Further, it is
very difficult todefine acriterion for optimalityand, therefore, the framework within which a numeric method can
be applied i s limited.
The inadequacyof numerical techniques is true for many
other functions needed in power system operations.
Human operators, when presented with problems, recognize the situation and decide on a course of action in an
all-encompassing way, implicitly taking account of many
factors. Thus we
believe that the best path to solve ill-structured power system operating problems is to construct a
knowledge-based program that emulates human operators.
WOLLENBERC AND SAKACUCHI: AI IN POWER SYSTEM OPERATIONS

The strength of AI techniques over conventional programming can be summarized a s follows:

1) flexibility: Expertsystemsare
suited to solving illstructured problems. Furthermore, the environment used
to construct expert systems allows them to be prototyped
rapidly and incrementally so that many alternatives and frequent updatings can be tried.
2) HighPerformance: Expert systems
try to implement the
level of performanceexhibited bya personwith recognized
expertise in the problem domain.
3) Understanding: Expert systems canexplain the lineof
reasoning used as well as the contents of the knowledge
base. This is a key element when the designer debugs the
system and increases the confidence of the user.
IV. DEVELOPMENT
OF AN AI APPLICATION

The development of AI programs usually involves different proceduresthan used in the development of numeric
programs. To illustrate this difference we review an application that deals with the diagnosis of power system faults
and is constructed using both numeric and AI programs.
The problem is to identify a faulty element in a power
system by observing the relay and circuit breaker tripping
signals. Fig. 3 illustrates a roughoutline for a conventional

I
r

CODING

I
Fi 3. The pipeline model

of software development.

program development sequence and is referred to as the

pipeline model. This term was chosen because conventional program development usually proceeds in a fairly
rigid sequence of steps. As with the restoration problem
discussed earlier, no algorithm exists for this problem. Initial efforts are made to design an algorithm for diagnosis.
A document, written inunexecutable natural language and
flow charts, is completed at the end of the first phase. The
following serves as such a sample document for system fault
diagnosis:
Fault diagnosis is made based on the relay and circuit
breaker tripping signals. First, deenergizedcircuit elements are grouped into several areas that are topologically
disconnected. Then, the following diagnostic algorithm is
applied to each area.
Determine protected elements o f all relays which operated in one area and calculate their conjunction. Three
cases are possible:
7) If a single element comprisesthe conjunction, then
assume the faulty element to be in that area.
2) If two or more elements are included in the conjunction assume that some relays have not operated
and designate one element within the area as the
probable faulty element. Check the consistency among
1681

Fig. 4. Powersystemfaultproblem.

theassumptions,
possible unoperated relays, and
measurements until the designated element is
accepted or eliminated. If eliminated try another element.
3) The third case involves no elements in the conjunction and results from more than double ,faults
within that area or some relay misoperations.Find a set
of active relays so that the set of protected elernents
has a nonempty conjunction. If successful, applry the
diagnosis algorithm to each set, otherwise assume relay
misoperation.
is begun. Thealgorithm might
Now coding the algorithm
be written ina system description language andits source
code compiled into Fortran source code. This is followed
by the usual compile, load, link, and run sequence which
requires repeated changesto the program source code and
may even require changes to the system description itself.
The human resources to accomplish this task are estimated to be eight man-monthsfor the first phase, four manmonthsforthesecond,andsixman-monthsforthelast.The
total amounts to one and a half man-years. Thefinal Fortran
source code i s estimated to be about 15 thousand lines and
will run in
about 10 son a 32-bit processcomputer to diagnose a fault as shown in Fig. 4.
The major human resource is needed during the design
and debugging of the algorithm. This is easily understood
sincethediagnosisalgorithm cannot bedescribed in a compact way. As it i s an ill-structured problem, one could not
finish a design document which specified the algorithm
completely. This results in missing specifications, incorrect
code, and frequent modifications as the design becomes

1682

clearer. Thisproblem is made worsebythe fact that numeric

programming languages are less readableand less understandable when applied to knowledge-intensive problems.
The experience in applying an expert system to the same
problem is quite different. The software development paradigm is illustrated in Fig. Sand iscalled the"rugbymode1."

Fig. 5. The "rugby model" of AI software development.

As in the pipeline model shown in Fig. 3, there are design,

coding, and debug phases. However, they are strongly
interrelated in the rugby model and rigid sequencing of
phases is not intended. The output of the design phase is
not a document but a knowledge base that can beexecuted
"asis"onthecomputer.Inotherwords,thegoalofthefirst
phase is to prototype the algorithmas quickly as possible
(rapid prototyping). As it i s executable on a computer, both
user and developer can observe how it works at an early
stage of the project. If the prototype meets requirements,
the knowledge base is transported to the target machine

PROCEEDINGS OF THEIEEE,

VOL. 75, NO. 12, DECEMBER 1987

without changes. If it does not, some changeswill be made

during its transport to thetarget machine.
The actual human resource needed to apply expert system techniques to the fault identification problem
was three
quarters of a man-year. Most of this time was spent in the
first phase. A sample of the knowledge base written in the
OPS5 rule base language is shown in Fig. 6along witha natural language description of the rules.
(p rda-for,locJ,backup-protection
0
0

Prolog. However, it is unclear at this time whether these

languages arethe best to use in implementation of AI into
EMS systems. As discussed in [16], there are basically two
approaches that can be taken:
1) Implement AI programs in languages such asPascal
or C that contain many of the necessary language features
needed in AI programming. This is illustrated in Fig. 7(a).
The alternative illustrated in Fig. 7(b)would implement the
AI programs directly in Lisp on conventional EMS hardware. This approach suffers, however, in that such hardware is often not efficient at running programs written in
Lisp.

(protection

^relay-typo
dz
'pro-type

STANDARD EMSHARDWARE

local
<xl>
<yl>

-id
^location-at
'objaot

AI PROGRAY

<Z>

(protectia

'rrhy-type
^pro,type
'id
^location,at
'object

FORTRAN
WRITTEN I L M R W :

dz

local
<x2>
( <y2>

REAL TYE OS

<> <yl> }

<Z>

0
0

fault-information
'ar*_rrky
'bockup-rahy
dz
local-backup-protection
'spec
'fwlt-at

nil
<Z>

0
0

1
ubich mems:
"If theback-upr&y
<xl> of type dz ot the lo.oation <yl>
that protecta tho eI.nnnt <L> ia opuated d
theback-uprelay
<x2> of type dz at tbo location <y2>
that protecta tho elemnt <z> is operated and
<y2> h not q u a l to <yl>.
then tho fault h concluded at tho ohment <I> md
it is clawad by locd back-upprotoction."

Note:

The following
notation
is
assumed.
(1) All syntax f o l o w a the OPS5 language [l!]..
(2)
An
attribute preceded
is
by
the symbol
(3) Avariable
ia quoted
by
the
symbols
< ond

to

AI PRoGaUls

STANDARD EMS H A R D W A R E

".

>.

Fig. 6. Rule-based representation of diagnosis knowledge

and its meaning in natural English.

Although it is still early in the development, experience

in applying AI technology to the diagnosis process has
shown benefits to both utility companies and manufacturers. To utilities, the new software promises advanced
automation of power system operations never possible
before. To the manufacturer AI has great potential in help
ing to control the
costs of an ever-increasing demand for
more complex software.
V. IMPLEMENTING
AI IN ENERGY MANAGEMENT
SYSTEMS

Implementing AI in an EMS is more difficult than adding

a new application program written in an engineering language such as Fortran. AI programs are generally written
in special languages because
of the needs for symbolic processing and in some casesthese languagesrequire special
hardware.
Most AI development work is currently done in Lisp or

WOLLENBERG AND SAKAGUCHI: AI IN POWER SYSTEM OPERATIONS

2) Attach special hardware to the E M S that runs Lisp or

Prolog efficiently as illustrated in Fig. 7(c).
The decision is further made difficult by the fact that AI
development thrives best in a "prototyping" environment
which may not necessarily be the same environment as the
"delivery" environment.
Going with thefirst approach allows the AI programs to
interface directly with existing database, display, application, and communication software but requires rewriting
(or translating) the programs if the development language
is different from the delivery language.Thesecond
approach eliminates the rewriting ortranslating but adds
the problems of communications between two different
processors.

1683

VI. POSSIBLEAPPLICATIONS
OF AI
OPERATIONS

TO

POWERSYSTEM

References [16] and [ I 7 list various power system operations problems as potential areas for the application of AI
in energy managementsystems. Talukdar and Cardozo [ I 7
make a su bjective classificationof the various problem areas
according to one of four metrics: operating cost savings,
capital cost savings, improved quality of service, and generalizability. In this paper we makeour own predictionsas
to the problems most needing AI attention and these are
gathered into three groups labeled real-time control, operations planning, and operator training.
A. Real-Time Control Problems
7) Alarm Processing:The alarm processing problem is
really an extension of the diagnosis problem. When a serious disruption occurs on the power system, operators can
be overloaded with alarm messages. Because many of the
alarm messagesare redundant or present information
related to the same event the operators may havedifficulty
in understanding precisely what has happened. The useof
AI to intercept alarm messages and presenta concise diagnosis is now under active development in several organizations, see [2], [3], [18]-[20].
2) Switching Operations: Statistics show that about 40
percent of the tasks at a power system control center are
related to operations on circuit breakers andline switches.
Therefore, theautomation ofthesetasks should benefit system operators. One potential application is the automatic
generation of switching sequences. Some work has been
done on verification of the switching sequences,[21].
Another application is the identification and isolation of
faulted line sections as shown in [22] and 1231.
3) Voltage Control: Incorporation of static optimization
techniques such as an Optimal Power Flow (OPF) is common for new control centers which desire to control the
system voltage profile. However, the control actions recommended by the OPF do not take account of the future
load prediction or past history of control actions and may
proveverydifficultto implement since manyof thecontrols
require manual entry by the operator. Liu and Tomsovic[5]
address this problem.
4) Restoration Control: A large-scale blackout may happen on a power system, although quite infrequently. The
fact that blackouts happen infrequently makes the operators job that much harder because ofthe limited exposure
to solving the problem of restoring the system. As a result,
most control centers have restoration plans andattempt to
train operators in restoration using training simulators.
However, the number of possible ways to restore a power
can changedepending on the state
system is very large and
of critical components at the time the blackout occurs. TO
this end, a system which supports operators by giving them
timely guidance and provides them with a tool for shortterm operations planning is quite desirable. AS shown in
[I],
AI software is essential in constructing such a system.

B. Operations Planning

5) LoadFlowPlanning:Load flows are run by system

operators to determine effects of planned changes to the
system and to help the operator study appropriate alter-

1684

natives should equipment loading falloutside appropriate

operating limits. An intelligent and friendly interface to the
load flow program will help theoperator in setting up cases
to be run, interpreting the resultsof
solved cases, and especially in how to interpret results if the load flow fails to converge.Fujiwara et a/. [6] describe an early effort which
proved effective in developing an intelligent load flow
interface.
6) Unit Commitment: One of the problems encountered
in the use of unit commitment programs is the difficultyof
expressing all the constraints that operators must meet in
scheduling units. Present practice in many control centers
with unitcommitment programs is to runthe program and
then alter the resulting schedule to meet constraints not
included in the program. Mokhtari et a!. [24] describe an
expert system that was developed to aid operators in adjusting the input data to the unit commitment program so that
the resulting schedule meets all scheduling constraints in
an optimal way.
\
C. Operator Training
7) Personal Tutoring: Large-scale training simulators are
installed in power system control centers to enhance operatorsskills, [25]-[28]. One point of view states that the operator acquires theseskills through the efforts of classroom
instruction and over-the-shoulder adviceof a training
instructorwhile solvingdifficult operating problemson the
simulator. Another point of view adds the capability of having the training simulator provide custom-tailored advice
for a specific operator and a specific training situation. The
authors of[29] and [30]describe such training facilities using
AI techniques.
8) Scenario Building: Another aspect of operator training
is the need to provide adequately difficult training
scenario
cases for the training sessions. These scenarios need to be
made difficult enough and specific enough so that targeted
levelsof skill can be reached
in each aspectof power system
operation. Building such scenarioscanbequitedifficult for
training instructors and the research reported in [31]
describes an expert system to allow the instructorto build
a scenario given a specific level of difficulty for the training
exercise andthe type of problem that is to be presented to
the operator.

VII. CONCLUSIONS
The need for the application of AI technology to power
system operations has been analyzed. Some initial work
done in this area is reviewed to show how it differs from
and what its strengths are over conventional numeric programs. We believe that this growing technology will have
a significant impact on futureEMS design and will allow a
level of system automation unattainable with present techniques. However, severaltechnological barriers have to be
surmounted before this takes place.
The inference mechanisms that we foreseeoperating in
an EMS must perform at very high speed to be useful in a
real-time environment. In the short run, this high performancewill beaccomplished by improvements in thedesign
of software. In the long run, new computer architectures
will be developed that more closely match the needs for AI
programs. In addition, we recognize that the usefulness of
AI programs dependsstrongly on the qualityof the knowl-

PROCEEDINGS OF THE IEEE, VOL. 75,

NO.

12, DECEMBER 1987

edge base built into the system. Thus we must be able to

acquire knowledge efficiently to be able to take the best
of a knowladvantage ofAI.The automatic acquisition such
edge base is discussed in [32] but its state is still primitive
and knowledge engineers
skilled in this area are absolutely
necessary.
The knowledgeof how to run a power
system is resident
in the engineers and managers of the electric utility companieswhilethecomputertechnologytobuildanAlsystem
belongs to suppliers. Therefore, close cooperation and joint
development projectsareabsolutely necessary to reach
successful implementations.Finally,peoplehavingthe
multidisciplinary skills of power system engineering, computer science, and cognitive science must be trained for
such tasks.

REFERENCES

[19] C. DeMontravel, A real time expert system for alarm handling in the futureEDF regional controlcenters, ClGRE Study
Committee Rep. 39, Sept. 1986.
[20] F. Hein, Expert system using pattern recognition by
real time
signals, CICRE Study Committee Rep. 39, Sept. 1986.
[21] K. Matsumoto, T. Sakaguchi, and K. Uemura, Verification of
switching operationswith temporallogic, i n Proc. 8th Power
System Computation Conf., pp. 1096-1102, Aug. 1984.
[22] C. Fuki and J. Kawakami, An expert system for fault section
estimation using information from protectiverelays and circuit breakers, /E Trans. Power Delivery, vol. PWRD-1, no.
4, pp. 83-90, Oct. 1986.
[23] K. L. Tomsovic et a/., An expert system as a dispatchers aid
forthe isolationof line sectionfaults, i n Proc. 7986PESTransmission and DistributionConf., Sept. 1986.
[24] S. Mokhtari, J. Singh, and B. F. Wollenberg, A unit commitment expert system, i n Proc. 7987 Power lndustry ComputerApplications Conf. (Montreal, Que., Canada, May 1987).
[25] J. R. Latimer and R. D. Masiello, Designof a dispatcher training simulator system, in Proc. 7977lEE PowerlndustryComputer Applications Conf., pp. 87-92.
[26] I. Susumago etal., Development of a large-scale dispatcher
training simulator, /E Trans. Power App. Syst., vol. PWRS1, No. 2, pp. 67-75, May 1986.
[ 2 7 D. Magee et a/., A large two computer dispatcher training
simulator, / E Trans. PowerApp. Syst., vol. PAS-104, no. 6,
pp. 1433-1438, June 1985.
[28] T. E. Dyliacco et a/., Considerations in developing and utilizing operator training simulators, / Trans. Power App.
Syst., vol. PAS-102, no. 11, pp. 3672-3679, Nov. 1983.
[29] J.D. Hollan etal., STEAMER: An interactive inspectable simulation-based trainingsystem, A/ Mag., vol. 5, no. 2, pp. 1527, Summer 1984.
[30] B. Woolf et a/., Teaching a complex industrial process, in
Proc. AAAl Conf. (Philadelphia, PA, Aug. 1986).
[31] R. D. Masiello eta/., Review of the EPRl project on operator
training simulator implementation, presented
at the 1984
EEI Engineering Computer Forum, Atlanta, GA.
[32] R. S. Michalski,J. G. Carbonel1,andT. M. Mitchell, Ed. Machine
Learning. Palo Alto, CA: Tioga Publ. Co., 1983.

T. Sakaguchi and K. Matsumoto, Development of a knowledge-based system for powersystem restoration, /E Trans.
PowerApp. Syst., vol. PAS-102, no. 2, pp. 320-329, Feb. 1983.
B. F. Wollenberg, Feasibility study for an energy management system intelligent alarmprocessor, i n Proc. 7985Power
industry ComputerApplications Conf., pp. 249-254, May 1985.
L. R. Larson etal., The NSP intelligent alarm processor project, in Proc.7985 Minnesota Power Systems Conf. (Univ.
Minnesota, Minneapolis, Oct. 8-10, 1985).
A. Doi and K. Uemura, Knowledge-based operation guidance method for voltage and reactive power systems, presented at the IFAC Symp. on Planning and Operation ofElectric Energy Systems, July 1985.
C. C. Liu and K. Tomsovic, An expert system assisting decision-making of reactive power/voltage control,in Proc. 7985
Power lndustry Computer Applications Conf., pp. 242-248,
May 1985.
R. Fujiwara etal., An intelligent load flow engine for power
system planning, in Proc.7985 Power lndustry Computer
Applications Conf., pp. 236-241, May 1985.
T. Sakaguchietal., Design and developmentof human interface i n engineering workstation for power
system planning,
Bruce F. Wollenberg (Senior Member, IEEE)
presented at the lFAC/lFlP/lFORS/lEA 2nd Int. Conf. on Analwas born in Buffalo, NY. He received the
ysis, Design and Evaluation of Man-Machine Systems, Sept.
B.S. and M.Eng. degrees from Rensselaer
1985.
PolytechnicInstitute, Troy, NY, andthe
Y. Kohno etal., An intelligent supportsystem for powersysPh.D. degree from The University of Penntem planning, in Proc. 7985 lnt. Symp. on Circuits and Syssylvania, Philadelphia.
tems. Also i n /E Trans. Power Syst., vol. PWRS-1, no. 2, pp.
He has held positionsatLeeds and North67-75, May 1986.
rop Co., North Wales, PA, and Power TechR. L. Moore etal. A real time expert system for process connologies Inc., Schenectady, NY. He is prestrol, i n Proc. ?st / Conf. on Artificial Intelligence Applientlywith theEnergy ManagementSystems
cations, IEEE publ. CH 2107-1/84, pp. 519-576.
Division of Control Data Corporation, PlyT. E. Murphy, Developing expert system applications for
moth, MN. He
has served as an adjunct faculty member in Electhe
process control, in Proc. 7985 /SA Conf. (Philadelphia, PA).
tric Power Engineering Department at Rensselaer and in theElecP.A.Sachs
et a/., Escort-An expert system for complex
trical Engineering Departmentat the University of Minnesota. His
operations i n real time, Expert Syst., vol. 3, no. 1, Jan. 1986.
research interests include power system operations, large-scale
D. G. Cain, BWR shutdown analyzer using artificial inteloptimization of power system problems, and application of artiligence
techniques,
Rep. NR-4139-SR, Electric Power
ficial intelligence techniques t o energy management systems.
Research Inst., July 1985.
C. L. Forgy, OPS5 users manual, Tech. Rep. CMU-CS-81135, Dep. Comput. Sci., Carnegie-Mellon Univ., Pittsburgh,
PA, July 1981.
B. D. Clayton, ART programming tutorial, InferenceCorp.,
Toshiaki Sakaguchi (Member, IEEE) received
1984.
the B.S. and M S . degrees i n 1969 and 1971,
R. A. Kowalski, Logic for Problem Solving. Amsterdam, The
respectively, and the Ph.D. degree i n elecNetherlands: North-Holland/Elsevier, 1979.
trical engineering i n 1981, all from Kyoto
R. P. Shulte et a/., Artificial intelligence solutions t o power
University, Kyoto, Japan.
system operating problems, paper 86 SM 3354, presented
He joined Mitsubishi Electric Corporaat the 1986 IEEE Power Engineering Society Summer Meet.,
tionin 1971, and has beenworkingin
Mexico City, Mexico.
research and development on power sysS. N. Talukdar and E. Cardozo, Artificial intelligence techtems technology at the Central Research
niques for power system operations, Rep. EL-4323, Electric
Laboratory. He i s currently the Research
Power Research Inst., Jan. 1986.
is i n the
Managerthere. Hiscurrent interest
H. Kaninsono, Alarm handling by control and load dispatch- analysis of human heuristics and their use for management and
ing centers in Japan, ClGRE Study CommitteeRep. 39, Sept.
control of large-scale systems.
1986.
Dr. Sakaguchi is a member of the IEE of Japan.

WOLLENBERG ANDSAKAGUCHI:

AI IN POWERSYSTEM

OPERATIONS

1685