Զѧ
ACTA AUTOMATICA SINICA
1997 23 5 Vol.23 No.5 1997



ڶANNģ͵ĸϵͳʱԤ
ߡ塡˺
ժҪһֶANNṹļֵѵ㷨ַӦڸϵͳʱԤ.ַ߳ʱԤȡǿģ͵Ŀɿ.ģΪԽһڶANNģ͵ĸϵͳԤ.
ؼʡANNṹǰ磬ϵͳԤ.
LONG-TERM PREDICTION FOR COMPLEX SYSTEM BASED
ON MULTIPLE ANN MODELS
GAO FENGLI RENHOU
(Systems Engineering Institute, Xi'an Jiaotong University, Xi'an 710049)
AbstractAn extremum clustering approach for multiple artificial neural networks is developed in this paper. Then it is used for long-term prediction of complex dynamic system. With this method the long-term prediction precision and model robustness can be improved. This approach can be used for complex system predictive control using multiple ANNs.
Key wordsMultiple ANNs, multi-layer feedforward neural networks, prediction of nonli-near systems.

1
ANNṹɶԶ໥ӡЭͬõANNANNģɵϵͳ.һ͵Ԫϵͳڸ߲϶Խṹģ.ֲλģ黯֯ʽŲӡ߼Ϊ.
öANNṹķԵõ߶ȿɿԪϵͳ˼ANNо14.Ϊ˽Ӻıƽ⣬Ҳ˶ANN5ݣظANNıƽ.
2ANNṹ
ANNṹͼ1ʾڲͨ͵磺һʵĳӳϵ磬Э֮ϵſ.һһANNṹУж磬һſ.ͬ뵫ɸԶſԸֵļȨͣ

(1)
ySʾsֵ.ڶANNṹ˵ĽṹʽģMFNNRBFFLNCMACȶǿȡģֻҪǾõĶԷӳıƽ.Ϊʹ򻯣һɲǰʽʽ

(2)
ͳһ.ſͬӦÿһһ(1)ʽgSʾs.ſǷӳϵеĽɫ.ݵǰʱ̵ֵĳԭļȨֵԸѡ.ſԾANNṹĹ.

ͼ1ANNṹʾͼ
3ڼֵĶANNѵ㷨
3.1ֵ㷨
еĶANNṹʵַֻǶռо࣬ûпõĻֽķֲϵֳӳ䱾ļԻ.
ͨMFNNķ֪MFNNʵֵӳɸڵһ鱥Ȩ.ʹкʽƽMFNNԺԵҪ.һصʹøкǵֵģôڸƽʱڵķ򼸺ֵָ෴ڵ֮ǳ٣ƽȽ.
ΪõֻʽҪϽмֵ.ԴһּݶЧֲֵķԶ̺ĺϵмֵ.
Ϊ{(xi,yi),i=1,2,P},ֵ㷨£
1)óʼֵ.ȷKѡKʼ.
2)ֵڽϵȷиĹϵ.ڵi(xi,yi)xi-ckxi-cj,j=1,2,,KCkCjΪk͵jӼľģ(xi,yi)DkDkʾkĵӼ.
3)Ĺϵ¾ģֲֵʱԵk࣬(xi,yi)Dkѡ(xi,yi)Ϊµĵkľ.ȡоֵΪµľ.ֲͬСֵʱȡоСֵΪµľ.
4)Ƿ˱仯ǣ򷵻2.
5)ǷٽǷڡci-ck,(ţ0)鲢ڵ࣬Ȼ󷵻2㷨.
һ㷨ΪɶѵĻ֣ľֲֵģֲСֵģĺϲΪֵļٰڽԭ(㷨2)».
ֵANNṹ繦ֺܻЭͬǣ룬ͬʱеԺϵľ.ֽͨ˺ĸԣȷʵֵģȷļ壬ĿҲĸ.
3.2ANNṹļֵѵ㷨
ɴ˿ԵõڼֵĶANNѵ㷨
1)ֵ㷨л֣ѡʼK=/50༰鲢ԵõKֵck,k=1,2,,K.ڽԭȷDk
2)ʼANNkڸѵӼDkڣԳʼƽEk
3)ҳƽANNkӦѵӼDkڣĳѵ㷨ѵT(ȡT=50)
4)ǷEkСڹ涨ޣѵ򷵻3.
繤׶ΣֻԵǰֵڽԭȷ࣬ӦĸõֵԴΪֵ.
ΪMFNNһϵƽԣÿƽӳĺϵЩĹͬþͿԵõƽ.
4ڶANNṹĸϵͳʱԤ
ANNģУΪſڲͬ϶ϵͳ.ſɵǶϵͳģʽ࣬ǶϵͳһִġԵ.ǶԸģʽƣǶϵͳһֶ.ˣɽſѵѵֿǰ߲üֵ㷨߽.㷨ڶANNʱԤģжԸ㷨.
һʵʶ󣬿⵽һʵʵ롪{u(t),yp(t)t=1,2,,T}.ͨʵʵƣ൱ĳ̶ϷӳĵĶ̬.ݣ߽ĳʱԤģ.
ԲӵNARMAģΪʱԤģͣANNģ£

(3)
gSʾſĵsySʾsֵʽ2.
ѵʱ1PԤģͱʶָ꺯Ϊ

(4)
Цs(t)=yp(t)-ys(X(t),s)tʱ̵ĵsĹ.J(P)ӳ˸1PԤ.J(P)ĶǼٶÿһҪƽȻǶԣֻӦһʱԸζ(gs)ǲͬ.
֪ģʼֵym(-Ky+1),ym(-Ky+2),,ym(0)u(-Ku+1),,u(0),u(1),,u(P-1)yp(1),yp(2)yp(P).ʱԤģ͵ѵΪڸģͽṹ(3)£ȷģͲ=(T1,T1,,TS)Tʹʽ(4)ָ꺯С.
ڸѵ㷨ǰĿ꺯Ըĵ

(5)
гΪƫ(orderedderivatives)6ݣʾԱȿǸֿܴڵĸϺϵ֮õƫ.(5)ʽеƫʽӼõ

(6)
ԲȷΪ

(7)
⼸㹫ʽпԿǰʱģͲƫͨǰKyԲƫ.֮ǰʱ̵ͨģӣ򴫲繤Ĺȥʱ.ʱķ򴫲˼.ʽ(6)Ӧͨsķ򴫲.
ҪָǣڼֵANNģڣһͨſķͨ·.ſֵֵͨõģÿһʱgkֻһ1ȫ0ſǲɵ.ڼX(t)ֻ΢Сı仯ʱӰgkȡֵǰ£ΪgkĵΪ0.˶ԼֵĶANNģֻͨķ򴫲ͨſķ򴫲.
MFNNķ򴫲㷨ɵõµĶANNʱԤģ߱ʶ㷨.ڵڦ-1θ֮󣬼ڦȴȨֵ
1.ʼ
2.t=1,2,,P
a)򡢷򴫲.ſtʱ̸ȨgS(t)Ա(6)ʽ
b)ۼݶֵ
c)
d)ʽ(7)
3.ĳַ¸Ȩֵ.ѡBP͸¹ʽ.õǻڱ߶ȷĽBP㷨.
5ֵ
Ƚһ¶ANNģ뵥ANNģ͵Ĳ.Ƕ̬ϵͳ

(8)
ȽһԤģ.辫ҪΪMSE0.0001.ANNģͣڵɱ߶BPѵ㷨2056εﵽҪ.ڵ6ʼӵ19.ԶANNģͣüֵѵ㷨ȶռл֣õ6ֵģֱ𾭹568,543675,522,520,495ѵﵽҪڵĿֱΪ8,7,6,8,8,8.һԤģͽгʱԤֱӵ1040ʱ̿ʼ30Ԥ.һԤģ͵ĻϣٲñķгʱԤģ͵ıʶ.ԵANNģͣ50Ϊ0.4577֮½.ԶANNģͣ68εΪ0.0741.ANNģ͵ЧƱȽ.ͼ2(a)(b)ʾͼʵ߱ʾߣ߱ʾģԤ.


ͼ2ģԤԱ
6С
һֻڼֵĶANNṹѵANNṹӦڸϵͳʱԤ˻ڶANNṹĳʱԤģ͵ʽʶ㷨.ĶANNṹļֵѧϰԣԴԤȣھж༫ֵԵһิϵͳԤһֿеķ.ĵķΪһڶANNģ͵ĸӿϵͳԤϵͳĻ.
ߵλͨѧϵͳо710049
ο
[1]Jacobs R A, Jordon M l et al. Adaptive Mixtures of local experts. Neural Computation, 1991,3:7987.
[2]Cho S B, Kim J H. combining multiple neural networks by fuzzy integral for rocognition. IEEE Trans SMC, 1995, 25(1).
[3]Narendra K S, Levin A U. Regulation of nonlinear dynamical systems using multiple neural networks., In:Proc. of ACC, 1990, 16091614.
[4]Hrycej T.A Modular Architecture for Efficient Learning. Int'l Joint Conf. on Neural Networks, San Diego. 1990, 1:557562.
[5]Li Renhou, Gao Feng. Complex function approximatino based on multi-ANN approach. Chinese Journal of Systems Engineering and Electronics, 1995,6(2):2231.
[6]Werbos P J. Backpropagation through time:what is does and how to do it? Proc. of IEEE. 1990, 78:15501560.
ոڡ1996-01-29
