00292702

Upload: sumit-dhall

Post on 03-Apr-2018

213 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/28/2019 00292702

    1/7

    IEE 2nd International Conference on Advances in Power System Control, Operation and Management, December 1993, Hong Kong

    DEVELOPMENT OF A KNOWLEDGE-BASED SYSTEMTO IMPROVE POWER PLANT THERMAL EFFICIENCY

    Alvin C.R.Wong C.Y.Teo H.K. HoNanyang Technological UniversityNanyang AvenueSingapore2263

    GbstractA prototype knowledge-based system to monitorthe real-time thermal efficiency of a 250 M W steam plantis discribcd. The performance of the steam plant is detcr-mined by evaluating the boilcr and turbine efficiency basedon ASMWANSI performance test code. Various abnormalor inefficient operating conditions can be d ctcctcd, ana-lysed and diagnosed. The knowlcdgc acquired has bccnencoded in the form of engineering models and objects,

    rules and mathematical formulae. The prototype systemcan be interfaced to the power plant's data acquisitionsystem and wil l provide u seful operation guide and diag-nostic aid to the plant operator to improve thermal cffi-c ency.Keywords : nowledge-based systems,simulation, powerplant thermal efficiency, diagnosis.

    1. IntroductionThe need for real-time diagnosis of steam plant hasbeen recognised for many years. Current economic and

    social factors put suingent requirement on steam powerplants to be operated at high efficiency. Although mostpower plants are equipped with extensive d ah acquisitionand performanc e monitoring system, this large volume ofdata and alarms usually do not provide many mean s of in -telligent interpretation an d diagonsis of problems for prc-dicting corrective measures. Power plant operators have tocope with the d a y today operational problems and are facedwith several hundreds of measurement values. It is ratherdifficult for them to analyse all the data and evaluateaccurately the performance of the plant operation. Thus,there is a strong tendency towards the development of'expert system to m onitor the real-time thermal efficiencyand to advise operation for approriatc actions.Few of thcsc knowledge-based systems for timecritical processes are in daily operation because they do notinclude special real-time requirements. This paper reportsresearch into thereal-time complex do main of steam powerplant operation by developing a prototype system whichcombines the strength of frame-based representation and arule-based paradigm to provide process systems modeling,causal reasoning and temporal infercncing. The devclop-ment was motivated by an immediate need for a powerful

    and flexible rwl-time knowledge-based system for thepower generation industries to improve thermal efficiency.

    2, Svstem A r c h i t e c mThe prototype knowledge-based system is imple-mented using G 2 object-orientated real-time expert systemshcll [ 1,2]. The architecture of thc system is shown inFigure 1 . Know ledge is organiscd asobjects, rules and for-

    mulae segregated into eight workspaces : boiler drum,superheater, reheater, economiser, feedwater supply sys-tem, turbinc/condcn sor and fuel system. Each of theseworkspac es is a collection of process equipm ent related toit s plant area. For example, boiler-drum workspace con-tains objects such as boiler-drum, pump, valve, sensor-transmitter, control Icr-blocks and their associated displayan d graphs.

    The process equipment in each w orkspace are rep-resented by objecls. Each object has a table of atuibutes.An attribute table contains knowledge about the object.Figure 2shows the objcct definition offuel-oil-pump anda partol its attribute table. Itdcfinesall theclassattributes,including the appearance of i & icon, its name and othervalues of its variable. All classes of objects exists within ahierarchy ofclasses. Each class in the hierarchy in herits theattributes of its superior class. I f the process equipmentclass has an Inllow atu ibutc, then all sub-classes of processequipmcnt such as valve, pump inherit that Inflow attribute.Each class of objects is defined by an object definition thatspecifics the icon, the conncction and the attributes. Theobject definition isan abstraction of theobject. Theo bjectsare graphically linkcd by a connection. Figure 2showsseveral connections joining pumps and olherequipm ent. Insteam power plant, a connection might be a water-pip e ora stearn-pipc. Connecting posts are used to connect objectsacross workspace, indicating that endpoints of connectionson separate work spaccs arc actually joined .

    A Vuriuble is a class of object that receive v aluesfrom external deviccsorG 2 inference engine. I t also has anattribute table attached which allows the developer tochoose source of data,validity interval and othcr attributesconsistent with a real-time environment. Senor variablesare designcd for realo r simulated measurements. Figure 3shows the I;.DF-A-dischurge-I)ressure attribute associatedwith thc FDF-A objcct.

    135

  • 7/28/2019 00292702

    2/7

    Knowledge BaseBuilderL

    -

    1

    Foreign FunctionInterfaceISteamProperty

    I2 File InterfaceorG2 Standard Interfacethrough conventionaldata acquisitionsystemI User ItI User InterfacetKnowledge BaseMeta-Knowledge

    Diagnostic Knowledge

    Real-time Inference EngintInheritanceForward chainingBackward chaining

    Plant DataA c q u i s i t i o n

    J I-ct7lant SimulatorFigure 1 :Theschematic architecture

    2.1 Real-time Inference Eneinehe FDF-A-discharge-pressure of FDF-A in Figure3 is a quantitative variable and has the value 281.0, asshown by itsLast recorded value attribute. Variables haveVafidiry nterval attribute, the length of time that the Lastrecorded value attribute of a variable will remain valid.Also, variables have a Hisrory keeping specifcatio n attrib-ute that you can use to keep histories of values that changesover time. Variables can get values from a number ofsources such as the inferen ce engine, the built-in simulator,formula or external data files. The Datu server attribute ofa variable indicates the source from w hich the values for thevariable can be obtained.

    Rules contain an expert's know ledge about what to con-clude and how to respond to the given sets of conditions.Most rules are generic in form and they apply to a wholeclass of objects.A formula is an equation that provides a value for avariable. Tw o types of formulae, specific formulas andgeneric formulae are used in the knowledge base. A

    specific formula is a formula that applies to just onevariable. Specific formula is used to calculate the value fora variable if the Data server for the variable indicates thesource is from the inference engine as shown in this ex -ample. A generic formula applies to a whole class ofvariables. An example of a generic formula for any fueltank is " th e volume of any tank =the level of the tank *thearea of he rank". Generic formula is used to calculate thevolume of any vertical tank, if the variable that gives thevolume of the tank has a Data server attribute that indicatesinference engine and if no specific formula exists for thevariable.

    The real-time inference engine reasons ab out the cur-rent state of th e application, and comm unicates it with theoperator or initiates other activity based on the state it hasinferred. T he inference engine operates on the knowledgecontained in the knowledge base, the simulated values andvalues that it has received from sen sors and other externalsources. The inference engine has the following functions:scans the rulesat the rates associated with each rule, focuseson the key objects by trying rules that are ass ociated witheach object regularly, invokes rulesof aparticularcategoryfor a particular class of object and backward chains orforward chains to other rules to find the values.2. 2 Plant Data Acauisition

    Through G2 standard interface, the real time data forboiler and turbine can be obtained from the conventionaldata acquisition system. Howev er, during developme ntstage, the real time data is read from external files using G2files interface. T he complete set of measurem ent coveringwhole range of boiler tempcrature, boiler pressure, gascontents, fuel flow, turbine temperatre and pressure at eachinlet and outlet, turbine enthalpy and various stages ofturbine steam Ilow are stored in different files at threeloading points, namely 120M W , 200 MW and 250 MW.All the measurem ents at other loading level are generatedthrough linear regression from the three loading levels.2.3 Plant Simulator

    The prototype system has a sim ulator of a hypotheti-

    136

  • 7/28/2019 00292702

    3/7

    FUEL-OII.-PUMP, an obiect-definitionFUEL-SYSTEM-OVERVIEW] I

    FUEL--OIL-PUMP-A

    Class nameSuperior cla ssCapabilities and restrictionsClas s restrictionsInhertited attributesDefault settingsAttribute displaysStubs

    Attribute specific to class

    FUEL--OIL-DAY-TANKColorIcon description

    FUEL--OIL-HEATER

    fuel-oi l-pumpPumpn on enonenonenonenoneinheritedan nput fuel-line fuel-oil-inletan output fuel -line fuel-oil-outlcinheritedinherited

    located at left 23 ;located at right 8

    Figure 2 : Fuel-oil-pump object definition in fuel-system-overview workspacecal, dynamic steam power plant model built within itself.Access to the plant sim ulator can provide users with th e op -portunity to experience dynamic situations and to repro-duce various situations in order to review relevant events.The plant simulator emulates the process dynam ics of keycontrol variablessucha s flow, pressureand temp eratureforvalves, pumps, vessels, heat exchangers and other processequipment.

    Time-based simulation is possible due to time-basedrepresentation of data in the development software. Everyvariablein thesystem, including allobjectattributes, has anassociated validity interval which specifies how long thevalue remains valid before another value must be re-quested. An internal time base is used to time-stamp alldata, to react to expired validily intervals in variables, andto scan rules according to their specified scan intervals.When the validity interval for a variable has expired, anumber of options are pursued for obtaining a new value,depending on the data server specification. One option forobtaining a new data is to specify a value or evaluate aformula based on another objects value. For example, thevalue for the flow to an outlet valve of a tank can bespecified as theoutflow of the tank connected to it. Anotheroption to determine the value of the object is to use back-ward chaining or forward chaining. A third option is toobtain the value from an external source such as a sensorinterface or a data file.

    Both shallow and deep simulation knowledge areused. A shallow simulation is a collection of simpleheuristics that models t h e observed behaviour. In con-stract, a deep sim ulation models the principles underlyingobserved behaviour. These principlesare th e causal knowl-

    edge that drives surface behaviour. Shallow simulation inis constructed using experimental test data from a steampower plant operating at steady state conditions embeddedinto mathematical functions and equations. Deep simula-Lion model in is constructed using simulation formulaebased on physical principles such as (volume * pressure /temperature) and discrete-state simulation formulae. Anexample of the latter category of formula is state variable: ext value of the pressure-at-ouller of an y pump =thepressure-at-inlet of the pump +the delta-pressure of thepum p, with initial value 1.015.

    Faults can be simulated manually by triggering certainfault conditions. When fault condition is simulated, thesimulatorchanges the valueofthe relevantactual datasuchthat the deviation between the actual and the simulatedvalue exceed acertain threshold value. W hen this happens,fault effects may pro pagate through causal path and ruleswill be activated to diagnose the fault. The fault simulationis only used w hen the data is not actually coming from theplant itself, and is used for the purposes of off-line testing/validation of the knowledge base and for training/tutoringof plant operators.2.4 User Interface

    The user interface is a link between the user, theknowledge base and the simulator. It contains graphicalrepresentations of meters, dials, alarm indicators, trendgraphics, and some objects in the steam power plant. Theknowledg e system uses these graphical representations tosimulate a user control interface and LO elp explain eventsin the plant. Figure4 showsan xample of the user interfacewith a cause and effect table to highlight the established

    137

  • 7/28/2019 00292702

    4/7

    IAIR-GAS-SYSTEM-OVERVIEW]FDF-A fandischarge-press I 281 Oinlet-temp1 30.5

    .FDF-A-D19CHARGE-PRESS. a ciu anti tative-varia bleOptionsUse r restrictions noneNames A -F D F -D I S C H A R G E-P R I S SData type quantity

    do no t forward chain. breadlhfirst backward chain

    I Formula 23.8.- 1 . 4 * x +0FDF-A E.1'.Initial-calue noneLast recorded valueHistory keeping specificationValiditv interval suppli ed

    Simulation detailsInitial value for simulation defaultData server infcrence engineDefault update interval 1 second

    28 1 O valid indefinatelydo not keep history.~....~~.~ 1.012 * x * xno simulation formula yetI I FDP-E3I ir-outlet I 298.1 1I as-idet I 340.0 1 1 air-inlet I 34.2 I[ gas-outletl 116.8 I I discharge-press I 274.0 IL inlet-temp] 30.5 ]

    ~ ~

    Figure 3 :FDF-discharge-press quantitative variable in air-gas-system-overvicw workspacefaul to operators. The cause is identified by the colour of thecircle icon of that row w h e n i t changed from green to rcd. Inthis example, the cause for reheater fau lt established by theprototype system was too much excess air after the alarms"RH utlet temperature low", "S H temperature low" and"Oxygen level exceed 4%" were detected.

    3. IntelliFent M onitoring Know ledgeIntelligent monitoring knowledg e consists of two parts

    :process equipment validation and powcr plant efficiencycomputation. For example, a particular equipm ent valida-tion rules capture human expertisc relating to obscrved be-havoirs of equipme nt failure. These heuristic rules repre-sent temporal knowledge and are capable of expressing dy-namic relationship. The followin g rule operates on thepremise that in a control process, a sensor operating nor-mally will pick up noise generated by the process. Thisnoise will cause fluctuation to th e output of the sensor. Asensorthatdoesnotexhibitsuch characteristic over aperiodof time can be concluded to be faulty.

    for any sensor Sif the rate of change per second of the block-output of Sduring the lust 10 seconds 10% onened 1.I t I+xygen level < .2%]SI1 temp. l ow

    Causes 1 1 1 1 1Too much excess airInsufficient G m pening 0 0 0R H tube dirty

    0 NormalAlarms

    Figure 4 :Cause and effect table for rchcatcr fault

    138

  • 7/28/2019 00292702

    5/7

    I

    Iblack-smoke

    LBoilerAir Heating Problem1 I Boiler GasHealing Problem]~

    Figure 5 :Comb ustion faults hierarchyand pressures through the turbine, the turbine heat rate andthe efficiency of the turbine can be calculated according toASME/ANSI performance test code 6. [4] [ 5 ] The grossoverall efficiency can be obtained by the product of theboiler efficiency and turbine efficiency. Both models aredeveloped using th e performance test data of a 250 MWgenerating unit. To evaluate the plant's performance, th ecalculated real-time efficiency is compared with the valueobtained during special performance test as well as all thehistorical data. When ever the real-time calculated valuedeviates from the expected value, fault diagnosis will beconducted. A foreign function w ritten in C is interfaced toG 2 to evaluate thermod ynamic properties of w aier/steamsuch as specific volume, specific entropy and spccilicenthalpy. [7]

    4, DiaLrnostic Knowled=The approach that a human expert uses in a real-timesituation is to m aintain a peripheral awareness across thedomain, watching for performance exceptions, and thenfocusing on areas of interest. The proposed inference en-gine operates similarly. The inference en gine continuallyscans knowledge wh ich the expert has specified for pen-pheral awareness. Mefa-level diagnostic knowledge in theform of mefa-rulescontrols the usage of relevant parts ofthe know ledge base to ensure efficient and effective diag-nostic reasoning. Since th e diagnostic emphasis is todetect, analyse and d iagnose faulty and ine fficient operat-ing conditions, top level sympton s such as excessive smokeopacity and high exhaust gas temperature are selected forperipheral awareness. For example, i f high exhaust gastemperature is detected at the stack over a long period oftime, the inference engine uses meta-know ledge to deter-

    mine which knowledge to invoke, thus focusing on specificarea of interest. Diagnostic knowledge in the form of time -sensitivei f fh .cn ,whicheverand other rules enable the b asiccauses of detected problems to be determined.

    This section describes th e method of combustion faultdiagnosis based on an example of the diagnostic knowledgeand reasoning structure. A combustion fault is consideredto have occurred if the smoke opacity monitored exceedsthe simulated value by a certain percentage. Figure 5(based on [8]) shows the combustion fault hierarchy en-coded in the knowledg e base. Diagnosis propagates fromthe top-level fault down from the detection of smokepresent at the stack to thec aus eof fault. Afteracombustionfault is detected, the colour o l moke is used to indicate thepossible top-level faults such as incomplete combustionfaults, excessive air faults and contaminated fuel faults.Once a top-level fault is determined, further diagnosticrules are invoked to deduce the root cause of the fault.Figure 6demonstrates th e implementation of th e com-

    bustion fault rules with in various workspaces. The Roo fworkspace conlains mainly mcta-rules employed to controlthe usage of releva nt workspaces to ensure efficient and ef-fective diagnostic reasoning. M eta-rules in the Root work-space consist mainly of scaning rules and forw ard chainingrules. Somc examplcs of thc meta-rules that activates theincomplefe combustion workspace are listed below. Theinvoke-combustion-fault rule has a fixed scan interval andwhen activated, it invokes the other three meta-rules thathave the category of combusfion-faulf. Fault detection isbased on the level of deviation between the m u d alue andthe simulated value corresponding to the parameters andmeasurem ents of the operating plant. The source of thesimulufed value, which relates the knowledge of the ex-pected behavoirof the plant isdcrived from thequantitativemodels of the various components of the plant and theperformance test data acquired from a 250MW generatingunit. B y utilising th e fault simulation facility provided inthe plant simulator, th e operator can simulate combustionfaults by increasing the actual smoke opacity value offurnace-1 vi a the user in te rk c so that it is greater than thesimulated value by 30 per cent. The diagnostic process is

    139

    http://iffh.cn/http://iffh.cn/http://iffh.cn/
  • 7/28/2019 00292702

    6/7

    Fault Workspace

    R m t RulesWorkspace

    activatc-excessive-air-fault

    activatc-fuel-faultPortion of Roo1 Rules Workspace conraining

    Figure 6 : Implementation of combustion fault rulesimplemented according to the following rulesRule name : nvoke-combustion-faultif the smoke-opacity of urnace-1 >(1.3 * the simulatedvalue of the sm oke-opacity of uranc e-I) then the smoke-present of unit-I is trueand invoke com bustion-fault rulesand inform the operator that Excessivesmoke detected atstack of unit-1Rule name: activate-incomplete-combustion-faultif the black-smoke of combustion-fault is truethen activate the subworkspace of incomplete-combustion-faultsand inform the operator that Blacksmoke present at stackof Unit 1Rule name :activate-excessive-air-faulti the white-smoke of combustion-fault is truethen activate the subworkspace of excessive-air faultsand inform the operator that Whitesmoke present at stackof Unit IRule name :activate-fuel-faultif the black-and-white-smokeof combustion-fault is truethen activate the subworkspace of uel-faultsand inform the operator that Black and white smokepresent at stack of Unit I

    When the top-lev el fault is concluded to be incom pletecombu stion then meta-rules will activate only the Incom-plete-combustion workspace. When a workspace is acti-vated, rules within that work space is also activated. Thefollowing two rules reside in the incomplete-combustionwork space are used to conclude the hyp othesis that insuf-

    ficiently heated fuel as th e cause for incomplete combus-tion fml t.Rule Name :confirm-low-fuel-tempif the black-smoke of combustion fault is trueand the low-fuel- emp of unit-1 is truethen inform the operator that Insufficiently heated fue lcausing incomp lete combustion and thus black smokeRule Name :check-low-fuel-tempifthe t-out of FCV201A

  • 7/28/2019 00292702

    7/7

    air heating problem, boiler gas heating problem, improp-erly atomised fuel, too much air, insufficient fuel andcontaminated fuel.5. Conclusions

    The knowledge acquired for theprototype system hasbeen encoded in the form of engineering models andobjects, rules and mathemantical simulations and formu-lae. The underlying object-oriented approach to system de-velopment enables the creation of flexible and genericknow ledge structure. Based on this idea, the main diffi-culty of building complex know ledge-based systems, thatof know ledge elicitation, is simplified. The behavior ofprocess compon ents can be made generic and used withindifferent plant models. The real-time knowled ge-basedsystem described in this research also represents a depar-ture from static know ledge-based design, as the issues ofreal-time considerations and dynam ic behavior have beenaddressed. Meta-level diagnostic know ledge in h e formof meta-rules control the usage of relevant parts of theknowled ge base to ensure efficient and effective diagnosticreasoning in real-time situations.

    It was discovered that there are a lot of similaritiesbetween a steam power plant and other complex processplants like oil refineries and chemical processing plants.As mentioned before, the prototype system can detect,analyse and diagnose abnormal and inefficient operatingconditions such a s low boiler efficiency and high stacktemperature. Although each system may serve differentneeds,they usethesame monitoringanddiagnosticconcepts.Based on this principle, the developed work provides a firmfoundation upon which future real-time knowledge-basedsystems with sim ilar process kno wledge requirem ent canbe built.Acknowledgement

    The authors acknowledge th e contribution of shf fmemb ers at the Public Utilities Board (Singapore) in pro-viding useful information and conducted preliminary re-search and testing. The last two authors would like tothanks their past students Mr. C.C. Lim and Mr. L.C. Limfor their contribution in the modellingand efficiency calcu-lation of the power plant.Reference[11 "G2 Reference M anual Version 3.0".

    Gensym Corporation, Cambridge, Massachusetts,1992.

    "ASME/ANSI Power Test Code for Steam Gener-ating Units". PTC .1, American Society ofMechanical Engineers, Reaffirmed 1979.AMSE/AN SI Performance Test Code 6 , SteamTurbines, American Society of MechanicalEngineers, Reaffirmed 1982.AMSE/AN SI Performace Test Code 6A , AppendixA to Test Cod e for Steam Turbines, AmericanSociety of Mechanical Engineers, 1982.ASME/ANSI Performance Test Codes 6S,Simplified Procedures for Routine PerformanceTests for Steam Turbines, Ame rican Socity ofMechanical Engineers, Reaffirmed 1979.Mayhew Y .R. and Rogers G.F.C., Thermodynam icand Transpo rt Properties of Fluids, Oxford :Basil Blackwell, 1980.Janet L. F a h , Christine M. M itchell, T.Govindaraj,"An ICAI Architecture for Troubleshootingin Complex, Dynamic Systems", IEEE Trans. onSystems, Man and Cy bernetics, Vol. 20, No. 3,pp. 43 - 547, MayIJune 1990.Duncan Rowan and Richard Taylor, "On-line FaultDiagnosis:The FALC ON Project", Artificial Intel-ligence Handbook, Vol. 2, pp. 379 - 399,Instrumention Society of America, 1988.Wendy B. R auch-Hindin, "A Guide to CommercialArtil'ical Intelligence", pp. 344-36 2, Prentice-Hall,1988.

    [2 ] Robert Moore et al, "The G2 Real-Time ExpertSystem", ISA Paper #88- 1627, InstrumentionSociety of America, 1988.

    141