T his article presents ASDEP, an expert system forr Idesigning a power plant's electrical auxiliary system,
implemented to demonstrate the feasibility of AItechniques for solving specific power system design
problems. Overall, ASDEP received positive commentsduring its evaluation. Specifically, Al techniques provedfeasible to address the particular design problem discussed.
Most design problems encountered in the electricalutility industry are complex and time consuming. Designersmust deal with interactions between alternatives,uncertainties, and important nonfunancial parameters inmost design phases. Typically, designers face thefollowing issues:
* Multiple objective functions;* Multiple (and often conflicting) constraints;* Complex system interactions;* The need for accuracy;* The need for trade-off, optimization, and "what-if'analyses; and
* Coordination of the decision-making process.Presently, human experts solve design problems based
on experience, specialized knowledge, and engineeringjudgment, using the computer only for analysis. Conven-tional CAD programs build on the basic assumption thatfeasible solutions do indeed exist to solve designproblems; however, this premise does not always apply inreal-world design. Human experts frequently deal withdesigns not meeting all constraints and objectives imposedupon them.
Developers have successfully applied AI techniques tosome difficult engineering design problems, but little hasbeen applied to electrical power system design. In thisfield, the major design problem areas are generationplanning, transmission planning, substation design, andauxiliary-system design.
Generation planning. Generation planningdetermines the appropriate generation mix (such as nuclear,fossil, hydro, or solar) required to satisfy future customerneeds at minimum cost while providing adequate returns onstockholder investments. Customer needs encompassconcerns from reliability and quality of electrical supply toenvironmental, political, and geographical impacts. Theseconcerns often conflict, and are further complicated bynumerous constraints, uncertainties, and trends imposed onthe design.
Electric utility generation planning is a multiobjectiveproblem; generally, no single strategy exists to optimizeall objectives involved. Furthermore, many objectives areinherently difficult to quantify and are, therefore, notsuitable for combination into a single analytical functiond
Transmission planning. Transrnission planningdetermines how many additional power system transmissionfacilities are required to meet the design criteria atacceptable capital costs. Design criteria cover such issuesas reliability, security, dynamic stability,
0885-9000/87/0200-0056 S01.00O1987 IEEE56 IEEE EXPERT
Figure 1. Flowchart of design process.
and environmental impact. This problem is, amplified by itssize (that is, the combinatorial explosion of possiblechoices), the uncertainty of predicting the future, and howfar in the future to plan.
Substation design. Substations supply electricalenergy from high-voltage transmission systems to users atlower voltage levels. Electric loads fed by substations arefundamentally random in nature; therefore, we needprobabilistic measures to ascertain design quality. Duringdesign, many initial skeletal configurations can potentiallybe used that somehow must be ranked before selections canbe made.
Auxiliary-system design. Power plant electricalauxiliary systems supply electrical energy, either from theplant or from the transmission system, to electric loads(such as circulation coolant pumps or measurement andsecurity monitoring systems) required to operate the plant.An electric power generating plant can produce full poweronly if its auxiliaries are fully operational. When power islost to the auxiliaries, a generating unit may have to reducepower or shut down completely. Typically, auxiliary electricsystems cost five to 10 percent of total plant cost.2 Duringa large plant's long-term outage, due to the repair of majorauxiliary subsystem components, corresponding replacementenergy costs can exceed the total auxiliary system's capitalcosts; hence, the need to carefully design such auxiliarysystems.
Of the four major design problems described above,auxiliary-system design offers the least formidable challengeinsofar as Al techniques are concemed. Auxiliary-system
design does not present as great a dimensionality curse asthe other problems do; moreover, verification is fairlystraightforward to implement.
This article will address various Al techniques proposedand implemented to design auxiliary systems, discussASDEP's features, and present an evaluation of ASDEPcarried out by an independent engineering design group.While our research has established the feasibility of Altechniques for solving specific power system designproblems, additional work is clearly required to lift ASDEPto industrial design levels.
Present design methodology related to electrical auxiliarysystems evolves from engineers and technicians who, afterexamining a plant's load requirements, propose possibleconfigurations. They analyze each configuration, performingeconomic short-circuit, load-flow, and motor-start studies.Then, subjective evaluations are usually performed based oneach design's reliability, flexibility, maintainability, andexpandability. The designers select the design scheme bestmeeting design goals. After selecting the design scheme,they produce equipment specifications and obtain bids.Following equipment selection, they reexamine the finaldesign.
Figure I (excerpted from a preliminary draft of the IEEEP666 Standard) provides a task organization flowchart of thedesign process, indicating approximate times required tocomplete each step. No provisions are made for feedback, or
SPRING 1987 57
to answer various "what-if' questions, until the end of theprocess.
Past experience reflects the adequacy of utility industryauxiliary-system designs. However, the following areas needimprovement:
* Clearly stated design and system performance criteriaare often not properly formulated, in writing, beforethe design.
* Typically, individuals with little formal trainingperform power system analysis.
* Few analysis programs supporting auxiliary-systemdesign are available.
* Most "what-if' questions are not explored because theywould delay the design process (note the lack offeedback loops in Figure l's flowchart).
* Due to the large staffs involved, overall systemobjectives can become unclear and conflicting.
* Insufficient reference material exists to supportsystematic auxiliary-system design.
Can knowledge-based systems alleviate some of thesedeficiencies?
The goal. ASDEP seeks to design good (as opposed tooptimal, since optimality is awkward to define related toelectric power designs) electrical auxiliary systems meetingoperational requirements. We measure design quality againstoperational performance, reliability, maintainability,flexibility, expandability, and cost. Consequently, weconceive ASDEP to be truly design-oriented.
Objectives and limitations. ASDEP evaluates howstate-of-the-art AI techniques can be applied to powersystem design, using practical and physical examples.Primary issues are to (I) evaluate AI potential, and (2)
implement existing AI techniques. This research does notattempt to produce a problem-free expert system, but ratherseeks to evaluate the feasibility of using general AIframeworks for such designs.
Since our research focused on demonstrating feasibility,we imposed the following additional limitations on ASDEPto highlight major concerns during its construction:
* Functional boundaries of the auxiliary system to bedesigned (for example, the switchyard) must be clearlydefined beforehand.
* Low-voltage systems are not explicitly considered(that is, less than 1000 volts); however, low-voltageloads are implicitly considered as lumped loads atintermediate voltage levels.
* We designed ASDEP to handle only single-unit powerplants.
* ASDEP's final output, not intended as a detailedengineering drawing, describes the auxiliary system'sfundamental topology.
We based our research on existing AI techniques;therefore, ASDEP's evaluation phase was very important asit might have been determined that Al techniques cannotadequately address problems related to power system design.We will specifically address evaluation in due course.
Observations incorporated in ASDEP. Trans-cribing problem-solving expertise from human experts tocomputer programs is at the heart of expert systemdevelopment.3 Design engineers rely on the followingobservations when designing auxiliary systems:
* Using little or no backtracking, they follow an orderlyprogression from initial designs (containing littledetailed information) to detailed designs.
* They use general problem-solving techniques in
conjunction with analytical methods during designdevelopment.
* They generally use analytical methods that arerelatively simple mathematical formulations relatingkey design parameters to targeted constraints.
* They cannot satisfy all constraints initially imposedon the design.
* They primarily seek satisfactory (as opposed tooptimal) solutions.
* They articulate preferences during the entire designprocess, called for by intermediate design results.
ASDEP embeds these six observations.
Let's consider ASDEP's general methodology, concen-trating on AI issues faced during its implementation. Latersections will briefly discuss an actual design and review anevaluation of an auxiliary-system design generated byASDEP.
Figure 2 illustrates ASDEP's organizational flowchart.Major features of the flowchart's basic blocks follow.
Block 1: Language processor. Designers use aproblem-oriented language, relying on a user-oriented andfixed vocabulary, to (1) conveniently input and display dataand (2) select and apply the appropriate rule group whenprompted by the program to do so.
Block 2: Input procedure. We must provideASDEP with specific design information before design canbegin. We sort this information into three parts: (1) loadand system data, (2) equipment inventory, and (3) goal andconstraint data. A major problem frequently encountered isthe lack of accurate information regarding load andequipment components. We use default rules in the know-ledge base to reasonably approximate missing load andequipment information, and can later modify this infor-mation as desired during design.
Block 3: Initial design procedure. The initialdesign proposal is an inherent part of most designprocesses. Fortunately, due to the nature of many powersystem design problems, a simple initial design can beproposed meeting some (but usually not all) design goalsand constraints. We can use two extreme methodologies toautomatically propose an initial auxiliary-system design, ahighly reliable but high-cost design, or a less reliable butlow-cost design.
Using the first approach, ASDEP would modify theinitial design by tearing it down to achieve lower cost whilestill meeting stated constraints. Using the second approach,ASDEP would modify the initial design by building it up, orby changing the proposed design's structure, to produce ahigher cost design meeting stated constraints. Humanexperts take an approach resembling the second method.
Figure 3. The initial design procedure.
ASDEP generates an initial design based on the secondapproach as well, intending to use top-down refinements toproduce a satisfactory design.
Figure 3 presents ASDEP's initial design procedureflowchart. The following additional comments areappropriate. The system
* Divides all loads into two broad groups, emergency (orvital) loads, and nonemergency (or nonvital loads),based on information obtained in Block 2;
* Creates the load buses simply based on load goals andon the related switchgear's kVA rating (buses are large,electrical conductors enabling electrical loads andsupply feeders to be connected);
* Distributes the loads on the buses, balancing the kVAloading on similar buses;
* Selects load bus voltage ratings based on rulescontained in the system's knowledge base;
* Determines the kVA load on each load bus by addingall loads connected to each bus; and
* Automatically proposes a radial distribution systemconnecting most load buses to the main generator andthe switchyard.
In an effort to remain at a preliminary stage, the initialdesign may not be fully connected. As will be seen, thedesign will be fully connected progressively as provided forin Block 6 (dealing with "critics") and in Block 7 (dealingwith design refinements). We use this approach, resemblingNOAH's underconstrained method,4 to avoid backtracking.
SPRING 1987 59
Figure 4. An example of default rules.
Figure 5. An example of a design change rule.
Block 4: Knowledge base. The knowledge base isan expert system's heart. KnYowledge acquisition is mostdifficult, frequently constituting a bottle-neck during expertsystem construction.3 How knowledge is acquired,represented, and organized represents a central issue for allexpert systems.
Most of ASDEP's application-specific knowledge isrepresented using production rules. ASDEP employs thefollowing rule language structure based on the Backus-Naurform:
::= (IF THEN )
where ANTECEDENT is a Boolean expression composed ofpreconditions that must be satisfied before a rule can beactivated, and where ACTION is a Boolean expression thatwill either change a design, select another rule group, orchange a design parameter.
We use a rule-based formalism to capture human expertknowledge and experience, and to store that expertise in theknowledge base. Knowledge base rules diverge into twomajor groups: equipment default rules, and design changerules. Whenever ASDEP calls the knowledge base, thesystem selects the appropriate rule group by usingmetarules.5 Next, ASDEP checks each rule's preconditionsand (if valid) performs the appropriate binding between rulevariables and the specific design. Due to the subjectivityinherendy attached to knowledge base rules, users must havethe option to add, delete, or modify all knowledge baserules. Figures 4 and 5 exemplify rules found in each majorgroup, using Lisp computer language code.
During any design's initial stages, users frequently lacksome information needed to start the design process.Typically, human experts supply their estimates of missinginformation based on knowledge and prior experience.Likewise, knowledge base default rules (intended to capturehuman knowledge and experience) will supply most missinginformation regarding load and equipment data.
Figure 4's first rule will supply the missing kVA motorrating, if required. In addition, a decision must be made forany design as to appropriate voltage levels; indeed, voltagelevel selection is a most significant factor in power plantauxiliawy-system design. Several factors affecting systemvoltage selections6 are
* Load magnitude,* Distance from the main power supply,* Utilization device availability (as a fu...