Page 50 - 《水资源与水工程学报》2023年第4期
P. 50
4 6 & ' ( ) & * + , - 2023 $
lished.Finally,theestablishedmodelsareusedtopredictandreconstructthecomponentsofpH,DO,
COD ,NH—N,meanwhile,WPT-GRU,WPT-LSTM andWPT-CSA-SVM,WPT-CO-SVM,
Mn 3
WPT-MGO-SVM modelsareestablishedforcomparativeanalysis.Thenthemodelsareverifiedbythe
predictionofpH ,DO,COD ,NH—NofGuanyinshansectioninKunming,YunnanProvince.There
3
Mn
sultsshowthatthepredictionaccuracyofWPT-CSA-GRU,WPT-CO-GRU,WPT-MGO-GRU,
WPT-CSA-LSTM,WPT-CO-LSTM,WPT-MGO-LSTM modelsforpH,DO,COD ,NH—Nof
Mn 3
thecasesectionarebetterthanthecomparisonmodelswithbetterpredictionperformance ,ofwhichWPT-
CSA-GRU ,WPT-CO-GRU,WPT-MGO-GRUmodelshavethehighestpredictionaccuracy;CSA,
COandMGOcaneffectivelytunethehyperparametersofGRUandLSTM,andsignificantlyimprovethe
predictionperformanceofGRUandLSTM;thesixproposedmodelshavehighpredictionaccuracyandsmall
calculationscale,whichcanberecognizedaseffectivepredictionmodelsforwaterqualitytimeseries.This
studycanprovidesomereferenceforrelevantwaterqualitypredictionresearch.
Keywords:waterqualityprediction;gatedrecurrentunit(GRU);longshortterm memorynetworks
(LSTM);waveletpackettransform (WPT);chameleonswarm algorithm (CSA);cheetahoptimization
(CO)algorithm;mountaingazelleoptimization(MGO)algorithm
[12] uh^?jY·1 ¶"8pBPænµ
1 ./01 [13]
(SSA-GA-BP) [qr; /Ë@ uh^U
uX@¡âfg[ÁZox@áW ·qB·z UÛ-qV8p 8ñNÉÊænµ
@8BFÕþÿ&) [1] 。£t, * MIKE21 (VMD-PSO-LSTM) [qr; ÃLA [14] uh
@¡qqrz, ÿ|@¡[^=pµ·? 3 [。 ^,°ÛqB·z ùò>Z23³(EEMD-
Ë 1 [´ÎÅ723?Ýï^"#[=p; Ë 2 SVR) [qr; OÈA [15] uh^U·qB·z
[´Î³I&ù?Ýï^[=p, Y BP(back BTænµ 8ñNÉÊænµ(VMD-CNN-
propagation) ænµ [2-3] 、 ùò>Z23³ [4] 、 8 LSTM ) x°@¡[qr。
jÛ, ²DIù×W3N´Q^'~,LSTM、
ñN É Ê æ n µ (longshortterm memorynet
works,LSTM) [5-6] 、 l#23ænµ [7] , + /Ôáa3(gatedrecurrentunit,GRU) £*¬
x@¡[ [15-16] ^;<¢。LSTM ´áæn
[=puXool^Qi, !´Æ<
µ(recurrentneuralnetwork,RNN) ^U½, /º?o
Æð|Æûù^°
¤Ão/[ÁZ^u
RNNÎV^@ZBHW@ZFC , ·
à, *tÞßg-V´Q8p}¶q, Ä9¥û
[ [17] 、 [;< [18-21] e¬R~。GRU´
¥^[.。Ë 3 [´x°[p, x°[p
LSTM^U½, K LSTM{º,GRU\ÿk¤^æç
·è½=¹, ¼½´'xèðèÎL[I^
ZÎ B k ¥ ^ [ l Q, * ¬ x v [ [ ;
x°[, Y TimeGAN-CNN(convolutionalneural [22-24]
< &¢。¾¬¢,GRU/LSTM õûù
networks)-LSTMqr [8] 、GRNN(generalregression
( ¨©ëù、 &ù 、 ÜÝëEÑù、dropout、Learn
[9]
neuralnetwork)-Markov qr 、STG(spatial-tem
RateDropPeriod ) o GRU/LSTM[lQþÿû
poralgraph )-LSTMqr [10] , x°[=p^
Iab, Rî3NêëÓ_óL¥å GRU/
Ñ´Jvqr^q<Ûº?a¼[qr^ê
LSTMõûù, êMX«ý3óþÿ^Óû
P, ![¯$$ûI^êÂ=l; ¼½´ Û, % ! Ø Ü Í Æ, I Î O P GRU/
'x·zH«^[=p, »=pJ·zH LSTM^[RX。*t, ³÷eý8p [18] 、 UÛ
«âfù×·z?bÏkþÇÈ^Ûfg·Z, -qV8p [25-28] -V´Q8p (swarm intelli
îoÛfg·Z}¶[、 &¶ô¥å±®[ gencealgorithms ,SIA) £q GRU/LSTMõûù
¯。Üx'x·zH«^[=pþÿû¥ ¢¥å¬, Ĥ¥oû¥^Óq.。
^[ÁZ, »=p£@¡âfg[;<¢ ?uX pHÆB COD 、DO、NH—NeZâ
3
Mn
[11]
¥åR~¬, Yø= uh^(·z 8ñ fg^[ÁZ, a} GRU/LSTM^[lQ,
NÉÊnµ(W -LSTM) âfg[qr; (õ> ~ SIA GRU/LSTMõûùÓq¢^¬Ö|, u