йѧ
CHINA ENVIRONMENTAL SCIENCE
1998 18 5 No.5 Vol.18 1998
Ƽڿ

BPRBF˹ԪԳ
̿ˮϵͳģıȽ* 
12ܡ11.ҵѧӦûѧϵ 150001
2.ѧ뻷ѧԺ 150001
ġժBPRBF˹Ԫ罨̿ϵͳģͣˮϵͳģӦԡоBPRBF˹Ԫĳ̿ϵͳģ׼ȷϵͳӰصĹϵ,ϵͳгľͶBP˹Ԫ罨ˮϵͳģ,,ƽٶȽRBF˹Ԫ罨ģ,ϲ,ƽٶȿ졣о˷ôͳģĲ㣬ΪʵˮϵͳŻṩ˿е; 
ؼʡBP˹Ԫ硡RBF˹Ԫ硡̿ϵͳģ
Comparative of BP and RBF artificial neural network's function on model-building of ozonation and biological activated carbon water purification system.Tian Yu1Wang Baozhen2Zhou Ding1 1.Department of Applied Chemistry, Harbin Institute of Technology,Harbin 1500012.School of Municipal and Environmental Engineering, Harbin University of Civil Engineering and Architecture, Harbin 150001.China Environmental Science.1998,18(5):394397
AbstractThrough setting up ozonation and biological activated carbon system model by BP and RBF artificial neural networks,the applicability of the two neural networks are investigated to the water purification system.The study shows that these models accurately describe the relationships among the influence factors of the system and economical ozone's dosage can be obtained comparatively though the model;the model established by BP network has a good general ability and a slow impending speed,on the other hand,the model established by RBF neural network has a bad general ability and a fast impending speed.The limitations of the traditional model-identification methods were get rid of.A means to realize the water purification system in line control is provided.
Key words:BP artificial neural networksRBF artificial neural networksozonation and biological activated carbon system model
ڳ̿ϵͳУˮˮʡͶʱ䡢̿ʱ䡢΢Ӱ໥á໥Ӱ죬ȷƽ׶ΪҪѡֵʱҪƵķΧֱӹϵϵͳˮˮʵĺûЧܵķӡ֮ܵԹϵϵͳȷԼ΢Ӱ죬̿ϵͳѧģǿհ,佨ģһϵͳ,ͳķӦϵͳھȺʵʱϵҪ󡣶Ϊһ庯ƽԪϵͳʶŻȴֳԽԡ1µؽ˹Ԫԭͷ뵽ˮϵͳķ,ͨʵֶBPRBF˹Ԫ̿ģϵͳơ桢,BPRBF˹ԪˮϵͳģӦԡΪʵָˮϵͳŻ,пˮƽ,ٽ乤ʵӦṩһ;
1BPԪRBFԪĽṹԭҪ
BPRBFΪо,ԪзչȽϳļ,ֱ˴಻֮ͬ2Ҫ:ԭϽ,BPȫֱƽ,RBFǾֲƽ磻ӽṹϽ,BPһ,ҲĲ,õĽڵ㺯ͨΪSigmoid RBFֻʽ,Ľڵ㺯ѡGaussͣѵϽ,BPǶǰʽ,һѵ,ԭݶ½ԭRBFѵ󶼲С˷ߵС˷3ӱƽϽRBFƽٶҪBP,BPķδ֪ԤҪRBF硣״ΰԪˮʵϵͳģ,ؿ߶ˮϵͳʵԺͷΧ,Խо塣
2̿ϵͳʵ
оCODMnֵΪˮʺûָ꣬ʵ1ʾ
1̿ϵͳвͬͶµCODMnȥ
Table 1Disposing rate of CODMn with different ozone amounts in ozonation biological activated carbon system

Ͷ
(mg/L)ԭˮCODMn
(mg/L)ˮ CODMn
(mg/L)ˮ CODMn
(mg/L)ˮ CODMn
ȥ(%)ˮ CODMn
ȥ(%)
1.004.494.221.936.0157.02
2.024.063.941.782.9656.16
3.034.844.211.6113.0266.74
3.724.274.151.652.8161.35
4.064.964.842.022.4259.27
5.014.564.521.980.8756.58
5.694.163.681.8414.8155.77
7.284.323.681.8414.8157.41
9.294.213.011.5428.5063.42
10.284.563.801.9616.6757.02
15.624.363.921.4410.0966.97

3BPRBF˹Ԫĳ̿ϵͳķʵ
Ϊѵʵ,ѡ˵BPRBF˹ԪĽṹͲȷһ㡢һBPԪڳ̿ϵͳĽģо,ϵͳȡѵǺܶ,ģģֻʹٶȼ,ˣмΪ1020,BPѧϰ㷨,Բñ׼BPѧϰ㷨Ŀѧϰ㷨,Ϊٶ,ڱовӦ䲽ѧϰ㷨,ԤڱƽԤȶΪ0.01ȷRBFṹʱ,ǵĶپƽǿ,ѵ,ѡȡɺƽ,ڱηʵо,ѡĽΧ610֮䡣ͬԭ,С㷨ȡȨֵ

ͼ1ʼBPԪģ
Fig.1Iniff.m initiated BP neural network pattern
 ȷBPRBFĽṹѵBPѵʵУʵݹһԹ뷽ʽ뵽Matlab,Neural Network ToolboxĺģInitff.m,Trainbp.m,Simuff.m,ʵѵ̡Initff.m,Trainbp.mģμͼ1ͼ2ڶRBFѵʱ,˳øģ,RBFѵͼ3ʾ

ͼ2Trainbp.m BP Ԫѵģ
Fig.2Training pattern of Trainbp.m BP neural network 
ģ˳ʵ˻BPRBF˹Ԫĳ̿ϵͳķ棬ͼ4ͼ5ʾÿͼְͼabcӣaпԿбƽSSESum Squared ErrorѵEpochӶٵ,վɲԺSSEС0.01,bѧϰ֮ͬĹϵ,ԿѧϰLRLearning Rateı仯԰ڶ,ָݲͬбʶȡֵķ,˼ٵãcѵԪʵ,ԭˮCODMnֵΪXᣬԳͶΪYᣬԳ̿ϵͳˮCODMn ֵΪZ

ͼ3RBFԪѵ
Fig.3Training circulation of RBF neural network 
˹Ԫάͼ˳̿ϵͳгͶˮˮCODMn֮ĹϵRBFԪͼBPԪģc)ͬ,ӳ˳̿ϵͳгͶϵͳCODMnȥʵӰ졣

ͼ4BPԪĳ̿ϵͳˮCODMnˮCODMnͶ
Fig.4The three-dimensional diagram of influent and effluent ,ozone amount in 
ozonation biological activiated carbon system on BP neural network
ɷ߿ԿBPRBF˹Ԫĳ̿ģ,ʵڼ϶Ʋѡȡ,ʡЧʣԸݻ˹Ԫĳ̿ϵͳģڸˮCODMnͳͶ,ˮCODMnֵҲڸˮCODMn,ʹˮCODMnﵽСĳͶû˹Ԫˮϵͳģ,ʹԤ⼰ڼ׼ȷ١ؽУԸģʾάͼʡ,кҪʵ,ӳˮһʱ,ͶԴЧӰ̶,ԵõȽϾõĳͶ,ȷ̿ϵͳĳȾкܴȽBPRBFķ,ԿڷӳϵͳӰعϵ,һµ,BPķͼֱ,RBFͼ𵴸,˵ڱηʵõݽṹ,BPʺҪ,RBFûоֲСƽٶȿƲûȫֳ

ͼ5RBFĳ̿ϵͳ
ˮCODMnˮCODMnͶ
Fig.5The three-dimensional diagram of influent and 
effluent,ozone amount in ozonation biological activiated
carbon system on RBF neural network
4
4.1˹ԪеķֲʽϢ洢д֯ѧϰص,˻BPRBF˹Ԫˮϵͳģ,˷ôͳģĲ㣬ø÷ͨϵͳݼɽΪ׼ȷģ,ģ͵ͨԼʱЧֻȡʵݵĹ㷺׼ȷ;ø÷ģͶڲͬˮ׼ȷԤ,ܹΪԱṩƲ,Ϊʵˮ߿ṩһе;  
4.2BP˹Ԫ罨ˮϵͳģ,õķ,ƽٶȽ,ھֲС,Ҫϵ͡ٵˮϵͳģͱʶ 
4.3RBF˹Ԫ罨ˮϵͳģ,ϲ,ƽٶȿ,ھֲС,Ҫϸߡ϶ˮϵͳģͱʶ
4.4BPRBF˹Ԫĳ̿ϵͳģ׼ȷϵͳиӰؼĹϵ,ԸݽģľͶ
ο
1Hush D R, Horne B G. Progress in supervised neural networks: what's new since Lippmann. IEEE Signal Processing Magazine, 1993, 10(1) : 839 
2Widrow B D.Learning in artificial neural networks a statistical perspeetive.Neural Computation,1989,1(4):425464
3Hornik K.Apporoximation capabilities of multilayer feedforward networks.Neural Networks,1991,(4):251257
߼
19684ʿ1997ҵѧʿվʿҪ³̿ϵͳȾЧоɺʡѧصĿ˹Ԫĳ̿ϵͳܽģоʹȵصˮȴĿ47ƪ
ոڣ1997-11-24
* ʡȻѧĿ
