Modified Predictive Optimal Control Using Neural Network-based Combined Model for Large-Scale Power Plants

Transcription

1 1 Modfed Predctve Optmal Control Usng Neural Networ-based Combned Model for Large-Scale Power Plants Kwang Y Lee, Fellow, IEEE, Jn S Heo, Jason A Hoffman, Sung-Ho Km, and Won-Hee Jung Abstract--Wth a Neural Networ-based Combned Model (NNCM) for a power plant, a Modfed Predctve Optmal Control (MPOC) system can be developed based on predctve control algorthms and ntellgent technques Durng the NNCM smulaton, an On-Lne Identfcaton (OLID) system s updated every few steps to provde nformaton from the model to the MPOC Moreover, the MPOC wll use the OLID as a test process to optmze the control actons, mnmzng tracng-error To search for the best control acton, the MPOC utlzes a heurstc optmzaton technque, Partcle Swam Optmzaton Wth the proposed MPOC system the only nput to the NNCM wll be the unt load demand Fnally, major outputs of NNCM wll be shown usng the proposed approaches, valdatng the procedure as a means to desgn a control system for a new power plant Index Terms-- Once-through type boler, super-crtcal boler, neural networs, modelng, modfed predctve optmal control, partcle swam optmzaton, power plant control, dstrbuted large-scale power plant T I INTRODUCTION HE development of large capacty power plants requres new approaches for control The fact that power plants are complex dynamc systems wth sgnfcant uncertantes, has led to a departure from conventonal control methods [1] Specfcally, for optmal power plant operaton, many control algorthms have been ntroduced usng ntellgent technques In general, the challenge n optmzng power plant operaton s to produce optmal control actons for mnmzng loadtracng errors The predctve control system can utlze an dentfcaton system to fnd the optmal control actons For both predctve control and dentfcaton, the classc control algorthms provde the means to desgn a control system wth a clearly defned mathematcal model However, whle power plants are gettng larger and more complex, conventonal control methods, whch mnmze quadratc objectve functons, have a large computatonal burden whch precludes generatng an optmal soluton n real-tme In conventonal PID control systems, a sngle logc falure K Y Lee, J S Heo, and J A Hoffman are wth the Department of Electrcal Engneerng, The Pennsylvana State Unversty, Unversty Par, PA (emal: wanglee@psuedu, juh138@psuedu, and jah1051@psuedu) S-H Km and W-H Jung are wth Corporate R&D Insttute, Doosan Heavy Industres and Constructon Company, Ltd (emal: and n a control loop can cause an entre system to go unstable Moreover, under a changng envronment, PID control systems have to properly tune gans and tme constants contnuously As an evoluton of conventonal PID algorthms, ntellgent control systems have been extensvely studed n recent years In order to produce optmal control acton, neural networ-based dentfers and heurstc optmzaton technques have been used for handlng nonmodel-based control system desgn and reducng the computatonal complexty n large-scale dstrbuted systems As a practcal approach, Model-based Predctve Controls (MPCs) whch ntally use open-loop optmal control, had attracted much attenton untl the ntroducton of Generalzed Predctve Control (GPC) The GPC algorthm uses closedloop optmal control wthn a movng horzon [2] However, both approaches requre a mathematcal model and much computatonal tme to fnd an optmal soluton By usng ntellgent technques such as neural networs, fuzzy logc, and evolutonary algorthms, an ntellgent predctve optmal control system was developed usng neuro-fuzzy dentfcaton [3]-[5] Wthout usng a mathematcal model, the ntellgent dentfcaton system, by updatng on-lne, can provde plant nformaton to the control system Although ntellgent predctve optmal control systems are desgned wthout usng models they requre a great deal of tme to fnd an optmal soluton In order to reduce the computaton tme, a new optmzaton technque s requred to be embedded nto the control system Ths paper presents Partcle Swarm Optmzaton (PSO) as a modern heurstc method for an onlne predctve optmal control system The PSO algorthm s based on the analogy of a floc of brds and a school of fsh [6] It has been well nown n many engneerng applcatons that PSO technques provde accurate solutons wth fast convergence and smple mplementaton [6]-[9] However, the performance of PSO n a predctve optmal control system s yet to be nvestgated Predctve control n general requres a repettve smulaton of the plant model However, n practce, ths s an mpossble tas for a large-scale plant Therefore, the Neural Networ Combned Model (NNCM) was developed to ease the smulaton The NNCM s made of a number of neural networ models, each representng subsystems of the plant Usng NNCM, a Modfed Predctve Optmal Control /07/$ IEEE

2 2 Fg 1 A 500 MW once-through type boler power plant (MPOC) system can be developed based on predctve control algorthms and ntellgent technques Durng the NNCM smulaton, an On-Lne Identfcaton (OLID) system s updated every few steps to provde nformaton from the model to the MPOC Moreover, the MPOC wll use the OLID system as a vrtual model to optmze the control actons, mnmzng tracng-error To search for the best control acton the MPOC utlzes a heurstc optmzaton technque, Partcle Swarm Optmzaton (PSO) Snce the NNCM requres external nputs other than the control acton, an External Neural Networ (ENN) s developed to provde the external data, such as the feedwater nputs and the sprays for the superheaters and reheaters Wth the proposed MPOC and ENN systems the only nput to the NNCM wll be the unt load demand Fnally, major outputs of NNCM wll be shown usng the proposed approaches Thus the proposed approaches provde the means to desgn a control system for a new power plant Followng the ntroducton, the 500 MW once-through type super-crtcal boler power plant s descrbed n Secton II Secton III descrbes the development of the Modfed Predctve Optmal Control systems Secton IV shows smulaton results to demonstrate the feasblty of the proposed approach The fnal secton draws some conclusons and presents future wors II ONCE-TROUGH TYPE BOILER POWER PLANT A Descrpton of Power Plant The power plant under nvestgaton s a 500 MW coalpulverzed once-through type boler-turbne-generator unt The controlled once-through type boler s capable of delverng steam at a pressure of 35 MPa and a temperature of 595 C Two forced draft fans supply ar to the burner and furnace, two prmary fans provde ar to the pulverzers, and two nduced draft fans are controlled to mantan furnace pressure at a desred value Economzers are arranged before and after a Selectve Catalytc Reducton (SCR) to mprove dentrfcaton and net effcency The superheater conssts of three parts, dvson, platen, and fnsh The steam s reheated after the Hgh Pressure (HP) turbne usng the prmary reheater and the reheater fnsh There s a separator on top of the furnace whch supples hgh pressure steam to the superheater dvson The waterwall s around furnace vertcally and sprally Flue gas s suppled to the furnace through the pulverzers and burners Fnally, the turbne generates power from the tandem compound trple turbnes, whch consst of three parts: a HP turbne, an Intermedate Pressure (IP) turbne, and Low Pressure (LP) turbne A depcton of the power plant s shown n Fg 1 Each subsystem nsde the furnace has common nputs and outputs: mass flow rate, temperature, pressure, and enthalpy of flud In addton to these nputs, there are control varables nvolved n drvng each subsystem to the desred state The model, whch s based on the ANN, uses the predefned control acton as feedforward control B Neural Networ-based Combned Power Plant Model Each subsystem of the boler depcted n Fg 1 was modeled usng a NN By combnng the models of the ndvdual subsystems the NN-based Combned Model (NNCM) was developed To reduce the complexty and provde better nformaton the prmary ar, forced draft fan, nduced draft fan, and ar preheater are clustered nto a sngle subsystem, called ar systems The waterwall and furnace are also clustered nto the furnace/waterwall subsystem The resultng sxteen subsystems wll be connected wth correspondng subsystem nputs and outputs; n addton, there are several external nputs for ar, water, coal, ol, and control actons Fg 2 shows the NNCM

3 3 Fg 2 NN-based Combned Model III DEVELOPMENT OF MODIFIED PREDICTIVE OPTIMAL CONTROL SYSTEMS A Modfed Predctve Optmal Control (MPOC) System General Predctve Optmal Control (POC) calculates a sequence of future control nputs on a predcton horzon [2] For large-scale power plants, the calculaton of a long range of the nput sequence demands an extraordnary amount of computatons and therefore requres a long tme to produce results The perod of predcton tme maes t dffcult to apply POC drectly n real-tme applcatons In a smlar manner the proposed MPOC system wll calculate future control nputs However, nstead of a long range of the nput sequence, the set of control nputs wll be produced only for the next tme step (one-step ahead predcton) Ths approach wll reduce calculaton tme to allow for real tme applcatons Fg 3 shows the overall structure of the control system From the Unt Load Demand (ULD) the MPOC calculates feedforward control acton and the External NN (ENN) system generates external nputs The feedforward control s Fg 3 Structure of the MPOC used by the On-Lne Identfcaton System (OLID) system as ntal canddate control nputs The OLID system estmates outputs from the control acton, and feeds them to the MPOC The MPOC then updates the control values and returns them to the OLID Ths process s repeated for a gven number of teratons before the optmal control acton and external nputs are sent to the NNCM B On-lne Identfcaton (OLID) System The OLID s used n the MPOC as a smplfed model of the NNCM to estmate the outputs that would be generated by a set of control nputs The OLID system s made up of 4 NNs whch use control values to estmate set-ponts the NNCM would produce Because the OLID has many fewer NNs and they are decoupled, the OLID system consderably reduces calculaton tme Morover, by usng the OLID as a test process for the MPOC, the OLID allows for the stablty of the NNCM to be preserved There are eleven stages n the NNCM where steam propertes are changed Each of these stages corresponds to the output of a dfferent subsystem The OLID approxmates the output of each of these subsystems usng control values as nputs For each subsystem the OLID provdes three set pont values: temperature, pressure, and enthalpy, for a total of thrty three set ponts There are thrteen control nputs used to drve the outputs of the plant to the desred set-ponts Ths gves a total of thrteen nputs and thrty three outputs Because of the large number of nputs and outputs, tranng a NN of ths sze requres a large amount of memory For ths reason the nputs and outputs were dvded nto four smaller networs that would requre less memory for tranng and produce more accurate results The order n whch the nputs and outputs are grouped s mportant The outputs must depend upon the nputs they are grouped wth, otherwse the

4 4 NN cannot accurately predct the correct outputs Table I shows how the control nputs and outputs are separated for the OLID Each output represents temperature, pressure, and enthalpy Before the OLID can be mplemented the NNs must be traned off-lne wth a wde range of data Data from the same ULD used to tran the NNCM subsystems s also used for off-lne tranng of the OLID networs TABLE I DIVISION OF INPUTS AND OUTPUTS FOR OLID Control Inputs Outputs On-lne ID#1 feedwater (u1,u2) feedwater On-lne ID#2 pulverzers (u3,u4,u5), economzer 1, nduced draft fan a (u6), economzer 2, furnace, forced draft fan (u7), nduced superheater dvson draft fan b (u8) On-lne ID#3 On-lne ID#4 prmary ar fan(u9), superheater platen spray (u10), fnal superheater spray (u11) hgh pressure turbne valve (u12), reheater spray (u13) superheater platen, superheater fnsh hgh pressure turbne, prmary reheater, fnal reheater, ntermedate pressure turbne When n run mode, the OLID system shown n Fg 4, uses the optmzed control nputs from the MPOC to generate estmated outputs The estmated outputs are compared wth the NNCM outputs and the error s fed bac to the OLID Usng ths error the weghts of each NN are updated every 5th data pont for more accurate results In search mode the estmated outputs are fed to the MPOC where canddate control actons are calculated and gven to the OLID The OLID returns a new set of estmated outputs to the MPOC, where the process contnues for the gven number of teratons When the MPOC has completed all teratons, the OLID returns to run mode, and the optmzed control actons are sent to the NNCM Fg 4 Bloc dagram of the On-lne Identfcaton System (OLID) C External Neural Networ (ENN) System The boler system of the power plant conssts of 28 nputs that orgnate from external sources The water provded to the feedwater system (Eu1, Eu2), the ar nputs to the ar gas system (Eu3, Eu4), the water for the reheater spray (Eu5, Eu6), and all control nputs are generated from other processes n the power plant The MPOC calculates the 13 control nputs, but another process s requred to provde the 15 external nputs to the NNCM It was determned that the external nputs were dependant on the ULD Therefore the external nputs to the NNCM can be accurately estmated usng NNs wth the ULD as the nput For ths secton a recurrent NN s also used because of ts ablty to capture the dynamc behavor of the nputs There are a total of sx non-constant external nputs whch are labeled Eu1 through Eu6 n Fg 5 To mprove the accuracy of the NNs the tme varyng external nputs were dvded nto three groups, each group correspondng to a sngle NN The frst NN has two outputs, the second has three outputs and the thrd has a sngle output The remanng external nputs are constant for the whole operatng range of the power plant They are smply suppled to the NNCM from pror nowledge of ther values Fg 5 Structure of the External NN system D Modfed Predctve Optmal Control (MPOC) The MPOC can be developed usng the proposed OLID, ENN, and NNCM Ths desgn wll preserve the stablty of the NNCM whle mnmzng the tracng-error Based on the ULD, the MPOC uses a mappng functon to fnd the feedforward control actons from a looup table The soluton space the MPOC uses to search for the optmal control nputs s then defned Multobjectve optmzaton by Partcle Swarm Optmzaton (PSO) s then used to fnd the control acton that wll produce the most desrable outputs Fg 6 shows the three procedures of the MPOC 1) Mappng Functon: Because the tranng ULD covers a wde range of operaton, the control nputs for the tranng ULD are used as a feedforward control for the MPOC To fnd the feedforward control nputs that correspond to the correct power output, a mappng functon that relates the new ULD wth the tranng ULD must be used The tranng ULD can be approxmated as a straght lne from 100% to 50% and therefore a lnear relatonshp can be used to estmate the feedforward control nputs based upon the power demand at

5 5 Fg 6 Bloc dagram of the MPOC any pont The mappng functon uses a new ULD to fnd feedforward control nputs correspondng to the same power output from the looup table The feedforward control nputs are then used as a startng pont for the PSO search 2) Soluton Space: Before PSO begns searchng for the optmal control nputs, a soluton space s needed as a boundary for the search In order to generate the soluton space, the control nput wndows use the feedforward control nput as the center of the soluton space or boundary gap The boundary gap, used to restrct the range of the soluton space, expands the search wndow as the ULD ncreases from 50% to 75%, and decreases the search wndow as the ULD moves from 75% to 100% The varyng lmtng functon restrans the optmzed controls from departng too far from the feedforward control, whch s the center of the soluton space A plot of the boundary functon s shown n Fg 7 The boundary gap sze for all control nputs s measured n percent In (1) Cost s the value whch s to be mnmzed, ranges from one to eleven because there are eleven subsystems for whch set-ponts are generated, a, b, and c are weghts that can be adjusted for multobjectve optmzaton, T, P, and H are nown set-ponts of subsystem for temperature, pressure, and enthalpy, respectvely, and, ET, EP, and EH are the respectve outputs of subsystem estmated by the OLID system By weghtng the dfferences between the set-ponts and the outputs of the OLID, PSO can smultaneously optmze the set of control nputs for multple crtera The PSO algorthm uses the cost functon (1) to update the control values whch are agan used as nputs to the OLID Ths procedure s repeated for the gven number of teratons before the optmal control values are used as nputs to the NNCM For the next tme step the ULD s used to fnd the Neural Networ Based Combned Model Soluton Space Cost * u,1 Optmzed Control Actons (gbest) Partcle =1 Partcle =2 * u, 2 Tme (sec) Partcle Swarm Optmzaton u, 1 u, 1 * u,13 u, 2 u, 13 u, 2 u, 13 Canddate Control Actons Gven Set-ponts + Partcle u, 1 u, 2 u =29, 13 Partcle u, 1 u, 2 u =30, 13 Estmated Outputs OLID =40 teratons Fg 8 Operaton of the PSO secton of the MPOC feedforward control and soluton space These parameters are gven to PSO and the process s repeated IV SIMULATION RESULTS Fg 7 Functon used to constran the soluton space for the PSO of the total range for that control nput 3) Multobjectve Optmzaton Usng PSO: The operaton of PSO for a sngle step of the NNCM s outlned n Fg 8 Usng the soluton space from the control nput wndows, PSO provdes canddate control nputs to the OLID The OLID estmates the outputs that would be generated by the canddate control nputs and feeds them to PSO The estmated outputs are compared wth the gven set-ponts and an error s calculated Ths error s used n the cost functon shown below, where some control nputs are weghted to gve them a hgher prorty than others: Cost = 11 = 0 a ( T ET ) + b ( P EP ) + c ( H EH ) (1) A Modfed Predctve Optmal Control (MPOC) For valdaton of Neural Networ-based Combned Model (NNCM), a new Unt Load Demand (ULD) s appled to the MPOC For valdaton purposes the new ULD should allow the plant to reach steady-state for several dfferent power demands Once steady-state has been reached the ULD should mantan that value for at least 30 mnutes A ULD wth these characterstcs replcates a typcal ULD that would be seen n real applcatons Fg 9 shows the ULD that was used to valdate the NNCM The ULD begns at 100% and decreases to 65% then ncreases to 80% and fnally bac to 100% of the MGR In the decreasng secton n Fg 9, the ULD s changng at a rate of 3 MW/mnute, n the ncreasng sectons the rate of change s 2 MW/mnute

6 6 The results of the MPOC smulaton for the new ULD are shown n the followng subsectons The four subsectons correspond to the four processes nvolved n the MPOC smulaton: On-lne dentfcaton (OLID), External Neural Networ (ENN), Modfed Predctve Optmal Control the range s between -1 and 1 By observng the dfferences between the estmated outputs of OLID and the outputs of NNCM, the accuracy of the OLID can be seen The OLID generates the estmated outputs very well by contnuous updates 2) External Neural Networs (ENN): The water temperature output of the ENN for the prmary reheater spray s shown n Fg 11 By comparng outputs of the ENN wth fltered data from the plant smulator, the ENN generates Fg 9 New ULD used for valdaton of the NNCM (MPOC), and Neural Networ-based Combned Model (NNCM) 1) On-lne Identfcaton (OLID): Fg 10 shows selected outputs of the OLID system Fg 10 (a) shows the superheater fnsh steam temperature Fg 10 (b) shows the prmary reheater steam pressure The OLID data s prescaled so that (a) Superheater fnsh steam temperature (b) Prmary reheater steam pressure Fg 10 On-lne dentfcaton result and target values for two outputs Fg 11 Water temperature for prmary reheater spray external nputs approprately to produce power and operate the NNCM smulaton 3) Modfed Predctve Optmal Control (MPOC): Sample control nputs, found usng PSO, are shown n Fg 12 Fg 12 (a) shows the prmary reheater spray control, and Fg 12 (b) shows the superheater platen spray control It s obvous that the MPOC generates control nputs that are qute dfferent from the fltered control nput from the plant smulator 4) Neural Networ Combned Model (NNCM) wth MPOC: The results from major outputs of the NNCM are shown n Fg 13: the total power output, the superheater fnsh steam temperature, the fnal reheater steam temperature, and the dvson superheater steam pressure Observng the error between the smulaton data and the NNCM output gves a good dea of the accuracy of the MPOC The prmary goal of followng the ULD has been acheved by the MPOC, whle mantanng the secondary objectves of eepng the pressure and temperatures wthn the operatng wndows Wth the proposed Neural-Networ based combned model, the Modfed Predctve Optmal Control system generates the optmal control acton very effcently Snce the PID-based plant smulator has many control loops, the falure of a sngle loop could adversely effect the operaton of the rest of the system Moreover, under a changng envronment, PID control systems have to properly tune gans and tme constants contnuously However, the MPOC generates optmal control acton by mnmzng load-tracng error wthout consderaton of complcated control loops or adjustng of gans and tme constants The MPOC also preserves stablty by predctng the outputs of the plant usng the OLID, whle PID control uses the actual error from the real plant Therefore the MPOC can be used as an advanced

7 7 (a) Prmary reheater spray control (a) Total power output (b) Superheater fnsh steam temperature (b) Superheater platen spray control Fg 12 Superheater platen spray control control methodology for use n real applcatons V CONCLUSION Usng the NN-based Combned Model (NNCM), a Modfed Predctve Optmal Control (MPOC) system s developed The MPOC conssts of an On-lne Identfcaton (OLID) system, and an External Neural Networ (ENN) The On-lne Identfcaton (OLID) system allows stablty to be preserved durng the search for optmal control nputs The External Neural Networ (ENN) provdes external nputs to the NNCM usng the gven unt load demand The proposed approach provdes a means to effcently search for optmal control nputs durng on-lne operaton The smulaton results show that the proposed MPOC follows the set-ponts and power outputs, and therefore can be appled n real tme to large scale power plants For future wor, wth the developed NNCM and MPOC, a reference governor wll be developed to provde optmal setponts and optmal feedforward control nputs for the MPOC Moreover, applcablty to larger capacty power plants such as an Ultra Super Crtcal (USC) boler power plant wll be nvestgated (c) Fnal reheater steam temperature (c) Superheater dvson steam pressure Fg 13 Results of the NNCM compared wth data from smulator

8 8 VI ACKNOWLEDGEMENT Ths research wor has been performed under a project awarded by Doosan Heavy Industres & Constructon Company, Ltd, Seoul, Korea The authors are pleased to express ther apprecaton to Dr Chang-Ho Cho of the Corporate R&D Insttute of Doosan for hs support of the project VII REFERENCES [1] K Y Lee, (M EL-Sharaw and D Nebur, Edtors), Chapter 12, Control of Power Systems, Tutoral on Artfcal Neural Networs wth Applcatons to Power Systems, IEEE Power Engneerng Socety, Publcaton #96TP112-0, pp ,1996 [2] K Y Lee and M A El-Sharaw (Edtors), Tutoral on Modern Heurstc Optmzaton Technques wth Applcatons to Power Systems, IEEE Power Engneerng Socety, IEEE Catalog Number 02TP160, Pscataway, NJ, 2002 [3] H Ghezelayagh and K Y Lee, Intellgent Predctve Control of A Power Plant wth Evolutonary Programmng Optmzer and Neuro-Fuzzy Identfer, Proc 2002 Congress on Evolutonary Computaton, vol 2, pp [4] H Ghezelayagh and KY Lee, Neuro-Fuzzy Identfer of a Boler System, Engneerng Intellgent Systems, vol 4, pp , 1999 [5] H Ghezelayagh and K Y Lee, Applcaton of Self-Organzed Neuro- Fuzzy Identfer n Intellgent Predctve Control of a Power Plant, Engneerng Intellgent Systems, vol 13, no 2, pp , 2005 [6] J S Heo and K Y Lee, Mult-agent system-based ntellgent steadystate model for a power plant, Proc the 13th Internatonal Conference on Intellgent Systems Applcaton to Power Systems (ISAP05), Washngton, DC, 2005 [7] J Kennedy and R Eberhart, Partcle swarm optmzaton, Proc 1995 IEEE Int Conf Neural Networs, vol IV, pp [8] J-B Par, K-S Lee, J-R Shn, and K Y Lee, A partcle swarm optmzaton for economc dspatch wth non-smooth cost functons, IEEE Trans Power Syst, vol 20, no 1, pp 34-42, Feb 2005 [9] C Reynolds, Flocs, herds, and schools: A dstrbuted behavoral model, 1987 Computer Graphcs, vol21, no 4, pp [10] K S Fu, Learnng Control Systems and Intellgent Control Systems: an Intersecton of Artfcal Intellgence and Automatc Control, IEEE Transactons on Automatc Control, vol 16, pp 70-72, 1971 [11] C C Ku and K Y Lee, Dagonal Recurrent Neural Networs for Dynamc System Control, IEEE Trans On Neural Networs, vol 6, no 1, pp , Jan 1995 [12] C C Ku, K Y Lee, and R M Edwards, Improved Nuclear Reactor Temperature Control usng Dagonal Recurrent Neural Networs, IEEE Trans on Nuclear Scence, vol 39, no 6, pp , Dec 1992 [13] J S Heo and K Y Lee "A mult-agent system-based ntellgent dentfcaton system for control and fault-dagnoss for a large-scale power plant," Proc IEEE Power Engneerng Socety General Meetng, Montréal, Québec Canada, 2006 [14] J S R Jang, ANFIS: Adaptve-networ-based fuzzy nference system, IEEE Trans on Systems, Man and Cybernetcs, vol 23, no 3, pp , 1993 [15] C-L-M Harnold, KY Lee, J H Lee and Y M Par, "Free-model based model reference adaptve nverse controller desgn for power plants," IEEE Power Engneerng Socety Summer Meetng, Edmonton, 1999, vol 2, pp [16] W-H Chen, D J Balance, and P J Gawthrop, Optmal control of nonlnear systems: a predctve control approach, Automatca, 39, pp , 2003 [17] S Kaneo, Hsatome, M; and M Hshda, Hgh effcency supercrtcal sldng pressure unts for ol/gas frng, Proc of Internatonal Conference on Energy Management and Power Delvery, vol 1, pp , 1995, Sngapore [18] P J Antsals and K M Passno, eds, An ntroducton to ntellgent and autonomous control, Kluwer Academc, MA, 1993 [19] H Branover, A El-Boher, E Greenspan, and A Bara, Promsng applcatons of the lqud metal MHD energy converson technology, Proc the 24th Intersocety Conference of Energy Converson Engneerng, vol 2, pp , 1989 Washngton, DC [20] T Inoue, H Tanguch, and Y Ieguch, A model of fossl fueled plant wth once-through boler for power system frequency smulaton studes, IEEE Trans on Power Systems, vol 15, no 4, pp BIOGRAPHIES Kwang Y Lee receved hs BS degree n Electrcal Engneerng from Seoul Natonal Unversty, Korea, n 1964, MS degree n Electrcal Engneerng from North Daota State Unversty, Fargo, n 1968, and PhD degree n System Scence from Mchgan State Unversty, East Lansng, n 1971 He has been wth Mchgan State, Oregon State, Unv of Houston, and the Pennsylvana State Unversty, where he s a Professor of Electrcal Engneerng and Drector of Power Systems Control Laboratory Hs nterests nclude power system control, operaton, plannng, and ntellgent system applcatons to power systems Dr Lee s a Fellow of IEEE, Assocate Edtor of IEEE Transactons on Neural Networs, and Edtor of IEEE Transactons on Energy Converson He s also a regstered Professonal Engneer Jn S Heo receved hs BS and MS degrees n Electroncs Engneerng from Inje Unversty, Korea, n 1999 and 2001, respectvely As a canddate, he s currently pursung the PhD degree n Electrcal Engneerng at the Pennsylvana State Unversty Hs nterests are multobjectve optmzaton n control systems, ntellgent dstrbuted control, multagents systems, modelng and control of fuel cell power plants, and real-tme embedded system Jason A Hoffman receved hs BS degree from Bucnell Unversty n 2005 He s currently pursung hs MS degree n electrcal engneerng from the Pennsylvana State Unversty Hs nterests are modelng and control of power plants, neural networs, ntellgent control systems, and alternatve energy systems Sung-Ho Km receved hs BS degree n Mechancal Engneerng from Yeungnam Unversty, Korea, n 1991, and MS degree n Mechancal Engneerng from Kyungpoo Natonal Unversty, Korea, n 1995 He has been wth Doosan Heavy Industres & Constructon, Co, Ltd, Korea, where he s a Prncpal Engneer of Plant Control System Team n the Corporate R&D Insttute Hs nterests are control, operaton and modelng of fossl power plant Won-Hee Jung receved hs BS and MS degrees n Mechancal Engneerng from Puyung Natonal Unversty, Korea, n 1995 and 1997, respectvely He has been wth Doosan Heavy Industres & Constructon, Co, Ltd, Korea, where he s a Senor Engneer of Plant Control System Team n the Corporate R&D Insttute Hs nterests are control algorthm, optmzaton, smulaton technque, artfcal ntellgence for fossl power plant