12 IEEE-final OnlineOptimalManagementofPEMFuelCellsUsingNeuralNetworks_看图王
See discussions, stats, and author profiles for this publication at: https://https://www.sodocs.net/doc/cf15739445.html,/publication/280087624
IEEE-final OnlineOptimalManagementofPEMFuelCellsUsingNeu Dataset · July 2015
READS
10
2 authors, including:
Ahmed Mohamed Azmy
Tanta University
111 PUBLICATIONS 390 CITATIONS
SEE PROFILE
All in-text references underlined in blue are linked to publications on ResearchGate, letting you access and read them immediately.Available from: Ahmed Mohamed Azmy
Retrieved on: 14 May 2016
Online Optimal Management of PEM Fuel Cells Using Neural Networks
Ahmed M.Azmy and István Erlich,Member,IEEE
Abstract—A novel two-phase approach to manage the daily operation of proton exchange membrane(PEM)fuel cells for residential applications is presented in this paper.Conventionally, the performance optimization is carried out of?ine since it is a time-consuming process and needs high computational capabil-ities.To simplify the management process and to enable online parameter updating,the paper suggests a new technique using arti?cial neural networks(ANNs).First,a database is extracted by performing of?ine optimization processes at different load demands and natural gas and electricity tariffs using a genetic algorithm(GA).Then,the obtained results are used for the of?ine training and testing of the ANN,which can be used onsite to de?ne the settings of the fuel cell.The tariffs and load demands as inputs of the ANN can be easily updated online to enable the ANN to estimate new optimal or quasioptimal set points after each variation in operating points.
The agreement between ANN decisions and optimal values as well as the achieved reduction in operating costs encourage the im-plementation of the proposed technique to achieve both fast on-line adaptation of settings and near optimal operating cost.This technique is applicable for different distributed generating units (DGUs),which are expected to spread within the power systems in the near future.
Index Terms—Genetic algorithm(GA),neural networks,oper-ation management,performance optimization,proton exchange membrane(PEM)fuel cells.
I.I NTRODUCTION
T HE reduction of energy cost produced in fuel cells is be-coming increasingly important in order to bring them to competition with combustion engines and other energy sources [1]–[3].The bene?ts behind utilizing fuel cells include high ef?-ciency,silent features,low emission,modularity,and the possi-bility of cogeneration applications.Among the different types of fuel cells,the proton exchange membrane(PEM)fuel cells have gained a lot of attention especially for residential applications [4].PEM fuel cells are candidates to be used in these?elds be-cause of their quick startup characteristics and high-power den-sities[5].
To decrease the generation costs in fuel cells,both the capital and the operating costs should be reduced.Optimal management of electrical and thermal power in fuel cells can signi?cantly contribute in achieving the required economical operation.To carry out this management,an accurate economic model has been developed in this study to describe the operating cost of
Manuscript received August29,2003;revised December3,2003.Paper no. TPWRD-00449-2003.
The authors are with the Institute of Electrical Power Systems,University of Duisburg-Essen,Duisburg47057,Germany(e-mail:azmy@uni-duisburg.de; erlich@uni-duisburg.de).
Digital Object Identi?er10.1109/TPWRD.2004.833893the unit taking into account both the electrical and thermal re-lations.Such a model is discontinuous and nonlinear in nature, and hence,necessitates the utilization of a suitable optimization tool[6].
The genetic algorithm(GA)has proven high capability for handling such optimization problems because of its?exibility and robustness[7]–[9].In this paper,the utilization of a real-coded multipopulation GA to optimize electrical and thermal power in PEM fuel cells is introduced based on the developed economic model.Obviously,the optimized settings are valid only for certain operating conditions and have to be recom-puted after each variation.The optimization process,which can be carried out only in the of?ine mode,is a complicated and a time-consuming task and,hence,requires high computational capabilities.However,it is expected that a large number of dis-tributed generating units(DGUs)will be utilized in the near future and,hence,it is important to standardize a simple man-agement method,which is probably adapted by the manufac-turer and has to be locally updated.Therefore,the paper sug-gests using GA only to provide suitable data for training and testing an arti?cial neural network(ANN)by optimizing the electrical and thermal power at different operating conditions. The problem of long execution time characterizing the GA and the need for repetitive optimization will be avoided due to the use of the ANN alone in the online mode.
The ANN is chosen due to its capability of learning and generalizing large scales of nonlinearities by extracting system features using training data[10]–[12].The implementation of ANNs is practically suitable for residential implementation due to its simplicity and low computational requirements.The ANN inputs in this application describe the current prices of fuel and electricity,the load demands in previous and present time inter-vals,and also the prognosis for a short time interval.To meet new requirements,modi?cations in the ANN parameters and also some structure changes,if necessary,can be accomplished in an easy manner(e.g.,over the Internet,by the manufacturer, and therefore,do not need advanced experience from the oper-ator).With the proposed method,the ANN will be able to adapt the optimal strategy to any variation in the operating conditions like tariffs and future power demands without carrying out a new time-consuming optimization procedure.
Analyzing the obtained results emphasize the signi?cant re-duction achieved in the daily operating costs using the manage-ment process,which contributes to improve the economic feasi-bility of fuel cells.At the same time,the agreement between the optimized values and the ANN outputs shows the effectiveness of the suggested method to provide a fast and simple manage-ment for the unit performance.The optimal management pre-
0885-8977/05$20.00?2005IEEE
sented in this paper is not limited to fuel cells;rather,it can be applied for many distributed energy sources.
II.E CONOMIC M ODEL OF PEM F UEL C ELLS
To achieve the optimal management of electrical and thermal power in PEM fuel cells,the daily operating cost“DOC(in dol-lars)”has to be minimized.The following equation describes the total daily cost of the fuel
cell:
(1)
where
DFC daily fuel cost for the fuel cell(in dollars);
DCPE daily cost of purchased electricity if the demand
exceeds the produced electrical power(in dollars);
DPSE daily income for sold electricity if the unit electrical
output power exceeds the electrical load demand(in
dollars);
DCPG daily cost of purchased gas for residential applica-
tions if the produced thermal power in the fuel cell
is not enough to meet the thermal load demand(in
dollars);
O&M daily operating and maintenance cost(in dollars);
STC daily startup cost(in dollars).
The daily fuel cost for the fuel cell is calculated as
follows:
(2)
where
natural gas price to supply the fuel cell(dollars per
kilowatt-hour);
time duration between two successive settings of the
fuel cell(in
hours);
net electrical power produced at interval J(in kilo-
watts);
power required for auxiliary devices(in
kilowatts);
cell ef?ciency at interval.
The ef?ciency of the fuel cell is known to depend on the gen-
erated active power and,therefore,a typical ef?ciency curve is
developed as a function of the electrical power[4]and used in
calculating the daily fuel cost.
It is assumed in this paper that the main grid system balances
the difference between the electrical load demand and the net
electrical output from the fuel cell.A daily cost has to be paid
for the purchased electricity when the electrical load demand
exceeds the produced electrical power.On the other hand,there
can be a daily income because of the sold electricity when the
electrical output power from the fuel cell exceeds the electrical
load demand.The following equations de?ne the above men-
tioned
terms:
(3)
(4)
where
and are the tariffs of purchased and sold elec-
tricity respectively(in dollars per kilowatt-hour)
and is the
electrical demand at
interval(in kilowatts).
Since the thermal output power from the fuel cell depends
on the electrical power,a nonlinear equation is developed to
describe this dependency[3].The daily cost of purchased nat-
ural gas for residential applications when the produced thermal
power in the fuel cell is not enough to meet the thermal load de-
mand is given
as
(5)
where is the natural gas price to supply residential loads
(in dollars per
kilowatt-hour),is the thermal demand at
interval J(in
kilkowatts),is the thermal power produced
at interval J(in kilowatts).
The fuel and electricity tariffs
(i.e.,,
and
)in the presented model represent the four decision
variables,which affect the optimal settings of the fuel cell.
The operating and maintenance cost are assumed as a con-
stant value per kilowatt-hour and,hence,it is calculated de-
pending on the produced energy.The startup cost depends on
the temperature of the unit and,hence,on the time terminated
before start up and is given as
follows:
(6)
where
and are the hot and cold start up costs,
respectively;
is the time duration,where the unit is off(in hours);
and
is the fuel-cell cooling time constant(in hours).
The minimization of the objective function(1)is restricted by
the following operational and technical constraints:
Unit capacity
constraint(7)
Unit ramp rate
constraint
(8)
Minimum up/down time limits(continuous running/stop time
constraint)
are
(9)
(10)
In the above
equations,
and are the minimum and
maximum limits of the generated
power
and are the
upper and lower limits of the ramp
rate,is the power gen-
erated at
interval are the unit on and off times,
MUT and MDT are the minimum up and down time limits,and
U is the unit on/off
status for running mode and0for
stop mode.
Finally,the daily number of starts and
stops
should not exceed a certain maximum
number
(11)
III.M ULTIPOPULATION R EAL-C ODED GA
The GA is a probabilistic search algorithm,which searches a
population of points in parallel.The real coding is used because
it provides better performance and faster conversion than the
binary-coded GA[7].The GA is implemented in this paper to
de?ne the optimal settings by minimizing the cost function(1)
subjected to the constraints given by(7)–(11).
AZMY AND ERLICH:ONLINE OPTIMAL MANAGEMENT OF PEM FUEL CELLS USING NEURAL NETWORKS 3
A.Constraints Representation
The economic model of the fuel cell is discontinuous and comprises many constraints.In case of tightly constrained problems,the infeasible solutions are known to cover the search space at the initial https://www.sodocs.net/doc/cf15739445.html,plete avoidance of infeasible solutions in this case leads to a high possibility of missing the area of global minimum [7].Also,moving the infeasible individuals to the nearest feasible area would be too complex and an extremely time-consuming process.Therefore,a penalty function approach is applied to convert the constrained problem to an unconstrained one by augmenting additional cost terms with the main cost function.The addi-tional terms assign nonlinear costs for solutions that do not satisfy constraints depending on their location relative to the feasibility boundary [7].
The adequate choice of the penalty functions and their pa-rameters is an essential factor in the evolution process.A higher additional cost value has to be assigned to any infeasible so-lution to ensure the rejection of all individuals that violate the constraints.Exponential penalty factors are used for ramp rate violation,while quadratic ones are applied when violating either minimum up/down time limits or maximum number of daily start-stop times constraints.This choice of penalty functions is found to achieve fast rejection of the infeasible members.B.Evolution Process
The ?owchart shown in Fig.1summarizes the main steps of the evolution process,where standard steps of the GA-based op-timization are extended using two additional processes.The ?rst is to apply the penalty function method to deal with the cur-rent constrained problem,while the second is to perform the migration between subpopulations.It is assumed that the oper-ating point of the fuel cell will be updated every 15min,which is the time duration “T ”as given in the economic model.The corresponding power reference values represent the optimiza-tion variables.For one day,this results in a total number
of
unknown variables associated with each individual.
At the beginning of the process,individuals are randomly ini-tialized in the power range between 0and 4kW to satisfy the ?rst constraint (7).
The evaluation of the individual performance is accomplished by calculating the total cost according to (1)and adding the cor-responding penalty term if it exists.The individuals are then ranked depending on their corresponding cost values and a suit-able ?tness value is assigned to each one according to its posi-tion within the population.To produce a new generation through the recombination process,strings with higher ?tness values are selected using the roulette wheel technique.The recombination process is based on two main steps:crossover and mutation.Crossover is applied to exchange information among the members of the population,possibly creating higher ?tting members.The max-min arithmetical crossover operator is used to carry out the crossover for the real-coded GA [7].Mutation is also an important recombination process to escape from possible local minima.The mutation in the real-coded repre-sentations is accomplished by disturbing the gene values with low probability.“Elitism ”technique is applied to ensure that
the
Fig.1.
Flowchart of the GA evolution process as used in the present research.
best solutions are not lost in moving from one generation to the next.According to this strategy,some of the ?ttest members of each generation are saved and copied into the next generation.The multipopulation structure is found to improve the quality of the results and,hence,it is applied in this study.The individ-uals migrate periodically between subpopulations to exchange information between them as shown in Fig.1.C.GA Parameters
Table I summarizes the GA parameters used in the optimiza-tion process.
D.Results of the Optimization Process
The GA optimization process is carried out at different elec-tricity and natural gas prices as well as at various daily load de-mands to provide a database for training and testing the ANN.This includes typical load curves at different seasons and real-istic tariffs.
Figs.2–5show the effect of varying the operating tariffs on the optimal operation of the fuel cell for a certain load demand.The base tariffs of the sold and purchased electricity,and natural gas to supply fuel-cell and residential applications are U.S.$0.07,U.S.$0.14,U.S.$0.04,and U.S.$0.07/kWh,respectively.Fig.2shows the effect of sold-electricity tariff
“
,”while Fig.3illustrates the effect of purchased-elec-tricity tariff
“
”on the unit operation.The effects of natural-gas tariffs for supplying both the fuel cell
“
”and
4IEEE TRANSACTIONS ON POWER DELIVERY
TABLE I
S UMMARY OF THE GA P
ARAMETERS
Fig.2.Effect of sold electricity tariff on the fuel-cell optimal operation.
the residential thermal load
“”are illustrated in Figs.4and 5,respectively.
The associated startup cost causes high operating costs in the case of repetitive start/stop cycles and,hence,individuals that lead to this situation are avoided during the evolution process.The priority is given for the continuous running or,in a few cases,for the complete shutdown of the unit.In some other cases,the optimal scenario is when the unit runs continuously for a certain time and then switched off for the rest of the day as the situations shown in Fig.3with the U.S.$0.12/kWh tariff and Fig.4with the U.S.$0.05/kWh tariff.In all cases,the last constraint (11)is always found to be inactive.
The results illustrated in Figs.2–5show the strong in ?uence of the four tariffs on the optimal settings of the units.In the liberal market,these tariffs vary frequently and,hence,the op-timization has to be often repeated,which will be a compli-cated and time-consuming process.The following section shows the implementation of an ANN to estimate the optimum,
or
Fig.3.Effect of purchase electricity tariff on the fuel-cell optimal
operation.
Fig.4.Effect of fuel tariff (for fuel cell)on the fuel-cell optimal operation.
near optimum,settings based on the results obtained by the GA optimization.
IV .M ANAGEMENT G ENERALIZATION U SING ANN During the past years,ANNs have gained wide success in many applications [10]–[13].ANNs provide a superior math-ematical tool for dealing with nonlinear problems due to their capability to learn and reconstruct nonlinear mappings.With a
AZMY AND ERLICH:ONLINE OPTIMAL MANAGEMENT OF PEM FUEL CELLS USING NEURAL NETWORKS
5
Fig.5.Effect of fuel tariff for thermal residential applications on the fuel-cell optimal operation.
suitable choice of the architecture and the corresponding param-eters,any continuous nonlinear relation can be approximated with appropriate accuracy.Once the ANN has been well trained, it will be able to be successfully applied to new situations,which are not used in the training phase.This generalization charac-teristic can provide solutions for many complicated problems in the power system in both online and of?ine modes.
In contrast to the GA optimization,which is a time-con-suming process and under some circumstances requires manual control,the calculation of the ANN outputs for a given input set is very fast and always de?nite.A further advantage of ANN is that it can be implemented on simple hardware structures as expected for residential fuel-cell applications.Therefore,the objective of this research is to use ANN to estimate the next optimal(quasioptimal)settings of fuel cells.It is obvious,that the ANN can only ful?ll this task when information from the previous operating points is available.In this way,the ANN considers the“past”and adopts the strategy for the future steps. Now,the optimization using GA is used to produce a suitable database for training and testing the ANN,which is carried out in the of?ine mode.After training and testing the ANN,it will be prepared for the online application.
A.Neural-Network Structure
Besides the input and the output layers,the network com-prises also three hidden layers,which are found to show accept-able accuracy.The number of neurons in the?rst,second,and third hidden layers are40,30,and20neurons,respectively.The tan-sigmoid transfer functions are used as activation
functions Fig.6.Neural-network structure.
for all of the hidden layers,while the log-sigmoid transfer func-tion is used for the output layer.The con?guration of the pro-posed ANN is illustrated in Fig.6.
The nodes in the input layer receive54inputs,while a single output is obtained at the output layer.The inputs to the net-work include two different natural gas tariffs for supplying the fuel cell and the thermal loads and two different tariffs for pur-chased and sold electricity.Also,information about the supplied thermal and electrical power at the present time interval,during the last three hours and in the next three hours is used as inputs to the ANN.According to our experiences,the incorporation of a time range
of h is suf?cient to achieve acceptable results. Also,the consideration of the number of startups and stops per day,which requires information throughout the whole day,is not necessary because this constraint was always inactive during the GA optimization.
The abovementioned structure of the ANN enables the sim-ulation of changes in the natural gas and electricity price even within the same day.Since the prognosis at the input layer is re-quired only for a short period,the error in forecasting the load demand will be small,and accurate results are expected to be obtained.
B.Training Process
The GA optimization process was used to produce the pat-terns for ANN training.Results corresponding to eight load curves were applied to train the ANN,while results regarding another daily load curve were kept for testing the trained net-work.For each of the load curves,different tariffs are used, which results in a total number of56278patterns that were gen-erated for training the ANN.
6IEEE TRANSACTIONS ON POWER
DELIVERY
https://www.sodocs.net/doc/cf15739445.html,parison between the GA-based optimal target and ANN output. The implementation of feedforward ANNs with a back-prop-agation learning algorithm has showed high capabilities in many applications[11],[12].In this paper,the online back-propaga-tion approach is followed,where the weights are updated after the representation of each pattern to the network.The chosen learning rate and error margin are0.01and0.05,respectively. The training is carried out for more than1000epochs,where the average root-mean-square(rms)error is less than0.06,which re?ects the accuracy of the training process.More than85%of the training patterns have rms errors less than the speci?ed0.05 error margin.
To evaluate the effectiveness of the training process,an ex-ample from the training results is shown in Fig.7.A compar-isons between the GA-based optimal target and the output from the ANN is given for200patterns.The conformity between the two curves re?ects the high accuracy of the training process,en-suring the high capability of the ANN to extract the features of the system.
C.Testing the Trained Neural Network
In the test phase,the trained ANN is activated by new testing data to verify the generalization capability of the neural net-work.One new electrical and the corresponding thermal load curves are used to test the trained network at different oper-ating conditions.This includes78different investigated cases all at the same load curve but at different natural gas and elec-tricity tariffs.A total number of5832patterns are used in the test phase,where more than65%of the testing patterns are found to have rms errors less than the0.05level.Fig.8shows the percentage difference between the daily operating cost using the ANN-based settings and the optimal settings as obtained from the GA-based optimization for the78investigated cases. Among the investigated cases,the ANN-based settings satisfy all constraints.
Figs.9–11illustrate results corresponding to three selected cases from the testing mode(i.e.,cases10,42,and63).Compar-isons between the GA-based optimal target and the output from the ANN are illustrated for the same load curve for three cases. The difference between the three cases is only in the values of the used tariffs.The ANN outputs show good agreement
with Fig.8.Percentage difference between the operating cost using ANN settings and the optimal
settings.
https://www.sodocs.net/doc/cf15739445.html,parison between the GA-based optimal target and ANN output (C;C;C,and C are0.03,0.05,0.16,and0.0$/kWh,respectively,“Case10”
).
https://www.sodocs.net/doc/cf15739445.html,parison between the GA-based optimal target and ANN output (C;C;C,and C are0.03,0.09,0.14,and0.1$/kWh,respectively,“Case42”).
the target,which demonstrates the capability of the ANN and also the effectiveness of the presented two-phase approach.
It is important to mention that the average optimal value of the total daily cost is about U.S.$3.689/day depending on an-alyzing the data used in the test phase.This value increases to
AZMY AND ERLICH:ONLINE OPTIMAL MANAGEMENT OF PEM FUEL CELLS USING NEURAL NETWORKS 7
TABLE II
C OST S A VINGS U SING THE P ROPOSE
D T WO -P HAS
E O PTIMIZATION A
PPROACH
https://www.sodocs.net/doc/cf15739445.html,parison between the GA-based optimal target and ANN output (C ;C ;C ,and C are 0.05,0.07,0.12,and 0.04$/kWh,respectively,“Case 63”).
U.S.$3.8691/day if the outputs from the ANN are used to de-?ne the settings for the fuel cell.This difference represents an inconsiderable value when compared with the achieved saving using the proposed energy-management technique.
To evaluate the effectiveness of the achieved reduction in the total daily cost,the results obtained using this strategy are com-pared with three conventional strategies of settings.The ?rst case is when the fuel cell operates always at its rated power,while the unit settings in the second case are adjusted to trace back the electrical load demand.The scenario in the third case is to follow the thermal load demand and to balance the differ-ence in electrical power through the electrical network.Table II gives the average cost in dollars per day as well as the av-erage difference of the three conventional settings and also of the ANN-based settings with respect to the GA-based optimal case.
The high operating cost of the ?rst conventional case indi-cates that this scenario is not relevant.The large produced en-ergy results in a surplus thermal power,which is wasted as no thermal storage is assumed.The unit in the second and third scenarios will cover either the electrical or the thermal load de-mand and,therefore,the cost is relatively decreased.Following the de ?ned quasi-optimal scenario,an average operating cost re-duction of U.S.$0.6131/day (U.S.$224/year)can be achieved compared with the best conventional settings,“the third case ”.This represents a signi ?cant reduction not only in the operating cost of the fuel cell but also in the total energy price.However,evaluating the approach depending on a comparison with one conventional case,the third one,is not fair and misleading.The type and natural of load demands affect the results considerably.
For other load demands,following the thermal demand can be the most expensive choice.
V .C ONCLUSION
In this paper,a new approach for optimizing the performance of PEM fuel cells for residential applications is presented.The main advantage of the proposed technique is its ability to pro-vide optimal settings without carrying out new optimizations after each variation in the operating conditions.A real-coded multipopulation GA is used of ?ine to carry out the optimiza-tion process for different operating conditions and load curves.Results from GA-based optimization are valid to be applied for the unit,however,they are used in this approach to provide a database for training and testing an ANN.The ANN is used to generalize the obtained results in order to avoid time-consuming and complex computations required by GA for repeating the optimization process.In this case,variations of load demands and operating conditions are simply simulated online by up-dating ANN inputs and,thereby,new quasioptimal set values are obtained.
The results show the capability of the proposed approach to achieve both the easy and fast adjustment of the fuel-cell set-tings and the reduction of the operating costs near the optimal values.The proposed technique can be implemented onsite with different DGUs to manage their performance when supplying residential loads.For onsite application,it will be enough to simulate the ANN structure through a suitable hardware device and use it online to accomplish the management process.
R EFERENCES
[1]M.W.Ellis,M.R.V on Spakovsky,and D.J.Nelson,“Fuel cell systems:
Ef ?cient,?exible energy conversion for the 21st century,”Proc.IEEE ,vol.89,no.12,pp.1808–1818,Dec.2001.
[2]M.C.Williams,“Fuel cells and the world energy future,”in Proc.Power
Eng.Soc.Summer Meeting ,vol.1,2001,p.725.
[3]I.Beausoleil-Morrison,D.Cuthbert,G.Deuchars,and G.McAlary,
“The simulation of fuel cell cogeneration systems within residential buildings,”in Proc.Bi-Annual Conf.Int.Building Performance Sim-ulation Assoc.–Canada ,Montreal,QC,Canada,Sept.11–13,2002.Session 1–6.
[4] F.Barbir and T.Gomez,“Ef ?ciency and economics of proton exchange
membrane (PEM)fuel cell,”Int.J.Hydrogen Energy ,vol.21,no.10,pp.891–901,1996.
[5] F.Ordubadi,“PEM fuel cells and future opportunities,”in Proc.Power
Eng.Soc.Summer Meeting ,vol.1,July 15–19,2001,pp.710–716.[6]Y .Zoka,H.Sasaki,J.Kubokawa,R.Yokoyama,and H.Tanaka,“An
optimal deployment of fuel cells in distribution systems by using ge-netic algorithms,”in Proc.IEEE Int.Conf.Evolutionary Computation,Volume 1,Nov.29–Dec.1,1995,p.479.
[7]I.G.Damousis,A.G.Bakirtzis,and P.S.Dokopoulos,“Network-con-strained economic dispatch using real-coded genetic algorithm,”IEEE Trans.Power Syst.,vol.18,pp.198–205,Feb.2003.
8IEEE TRANSACTIONS ON POWER DELIVERY
[8]P.Zhang and A.H.Coonick,“Coordinated synthesis of PSS parame-
ters in multi-machine power systems using the method of inequalities
applied to genetic algorithms,”IEEE Trans.Power Syst.,vol.15,pp.
811–816,May2000.
[9]S.Mishra,P.K.Dash,P.K.Hota,and M.Tripathy,“Genetically
optimized neuro-fuzzy IPFC for damping modal oscillations of power
system,”IEEE Trans.Power Syst.,vol.17,pp.1140–1147,Nov.2002.
[10]S.H.Ling,https://www.sodocs.net/doc/cf15739445.html,m,F.H.F.Leung,and Y.S.Lee,“A genetic algo-
rithm based fuzzy-tuned neural network,”in Proc.12th IEEE Int.Conf.
Fuzzy Systems,vol.1,May25–28,2003,pp.220–225.
[11]T.T.Chow,Z.Lin,C.L.Song,and G.Q.Zhang,“Applying neural
network and genetic algorithm in chiller system optimization,”in Proc.
7th Bi-Annual Conf.Int.Building Performance Simulation Assoc.,Rio
de Janeiro,Brazil,Aug.13–15,2001.
[12]W.Charytoniuk and M.-S.Chen,“Very short-term load forecasting
using arti?cial neural network,”IEEE Trans.Power Syst.,vol.15,pp.
263–268,Feb.2000.
[13]L.-H.Chen and C.-H.Chiang,“An intelligent control system based
on multiobjective genetic algorithms and fuzzy-neural network,”Proc.
IEEE Int.Conf.Syst.,Man,Cybern.,vol.3,p.6,Oct.6–9,
2002.
Ahmed M.Azmy was born in1968in El-Menoufya,
Egypt.He received the B.Sc.and M.Sc.degrees
in electrical engineering from the El-Menoufya
University,El-Menoufya,Egypt,in1991and1996,
respectively.
Currently,he is with the University Duis-
burg-Essen,Duisburg,Germany,supported by an
Egyptian scholarship for Ph.D.research.His Ph.D.
thesis focuses on the modeling and intelligent
management of interconnected decentralized power
generating units.From1992to2001,he was with the Power Engineering and Electrical Machines Department,Faculty of Engineering,University of Tanta,
Egypt.
István Erlich(M’99)was born in1953.He received
the Dipl.-Ing.degree in electrical engineering in1976
from the University of Dresden,Dresden,Germany,
where he received the Ph.D.degree in1983.
Currently,he is Professor and Head of the Institute
of Electrical Power Systems at the University of
Duisburg,Duisburg,Germany.He was in Hungary
in the?eld of electrical distribution networks.From
1979to1991,he was with the Department of Elec-
trical Power Systems of the University of Dresden.
From1991to1998,he was with the consulting company EAB in Berlin,Germany,and the Fraunhofer Institute IITB Dresden, Dresden,Germany,respectively.During this time,he also had a teaching assignment at the University of Dresden.His major scienti?c interest is focused on power system stability and control,modeling and simulation of power system dynamics including intelligent system applications.
Dr.Erlich is a member of VDE.