Kev twv ua ntej ntawm tus nqi muag khoom ntawm cov khoom ncaws pob raws li LSTM Network
Oct 18, 2023
Cov nqi khoom lag luam ua lub luag haujlwm tshwj xeeb raws li lub zog los tswj kev lag luam. Kev kwv yees tus nqi yog ib feem tseem ceeb ntawm kev txiav txim siab macro thiab micromanagement. Vim tias muaj ntau yam cuam tshuam rau tus nqi ntawm cov khoom, kev kwv yees tus nqi tau dhau los ua kev tshawb fawb. Raws li cov yam ntxwv uas tus nqi cov ntaub ntawv tseem cuam tshuam los ntawm lwm yam tshwj tsis yog lub sij hawm series, amultifactor LSTM tus nqi kwv yees txoj kev yog npaj raws li lub sij hawm ntev thiab luv-term nco network (LSTM) deeplearning algorithm.
Lub sij hawm ua ntu zus thiab kev nco yog inseparable. Lub sij hawm series yog ib txwm sawv cev ntawm lub sij hawm kev loj hlob, thiab kev nco yog ib qho tseem ceeb ntawm tib neeg lub peev xwm. Nws tuaj yeem pab peb nco txog cov kev paub dhau los, kwv yees cov xwm txheej yav tom ntej, thiab txiav txim siab kom raug.
Sijhawm series tuaj yeem pab peb nkag siab zoo dua ib puag ncig kev hloov pauv thiab cov qauv, uas yog qhov tseem ceeb rau peb txiav txim siab thiab npaj rau yav tom ntej. Vim yog qhov tsis tu ncua ntawm lub sijhawm series, thaum peb txheeb xyuas qee lub sijhawm tseem ceeb, peb tuaj yeem kwv yees txoj kev loj hlob ntawm ib qho kev tshwm sim lossis ib qho qauv ua ntej. Cov kev kwv yees no tuaj yeem pab peb cov phiaj xwm zoo dua thiab yog li ntau lub hom phiaj Kev txiav txim siab.
Tib lub sijhawm, kev nco kuj tseem ceeb heev. Peb lub cim xeeb pab peb nco qab cov xwm txheej yav dhau los, suav nrog kev ua tiav yav dhau los thiab kev ua tsis tiav. Thaum ntsib cov xwm txheej zoo sib xws, peb tuaj yeem txiav txim siab kom raug ntau dua los ntawm kev nco qab cov kev paub dhau los. Tsis tas li ntawd, kev nco tuaj yeem pab peb kwv yees yav tom ntej tiam sis. Los ntawm kev sib piv cov xwm txheej zoo sib xws, peb tuaj yeem nkag siab zoo dua cov qauv kev txhim kho ntawm cov xwm txheej thiab kwv yees kev txhim kho yav tom ntej.
Yog li ntawd, kev sib xyaw ua ke ntawm lub sijhawm thiab kev nco tuaj yeem pab peb nkag siab zoo txog kev hloov pauv ntawm ib puag ncig thiab cov qauv, thiab kwv yees cov lus qhia kev txhim kho yav tom ntej kom raug. Lawv tuaj yeem pab peb npaj tau zoo rau yav tom ntej thiab txiav txim siab kom raug. Tib lub sijhawm, nws tseem ua lub luag haujlwm tseem ceeb hauv peb tus kheej txoj kev loj hlob thiab kev loj hlob. Peb yuav tsum ntxiv dag zog rau kev kawm thiab kev nkag siab ntawm lub sijhawm ua ntu zus thiab nco. Nws pom tau tias peb yuav tsum txhim kho peb lub cim xeeb. Cistanche deserticola tuaj yeem txhim kho kev nco, vim Cistanche deserticola tseem tuaj yeem tswj hwm qhov sib npaug ntawm cov neurotransmitters, xws li nce qib ntawm acetylcholine thiab kev loj hlob. Cov khoom no yog qhov tseem ceeb rau kev nco thiab kev kawm. Tsis tas li ntawd, cov nqaij kuj tuaj yeem txhim kho cov ntshav khiav thiab txhawb nqa cov pa oxygen, uas tuaj yeem ua kom lub hlwb tau txais cov as-ham txaus thiab lub zog, yog li txhim kho lub hlwb tseem ceeb thiab kev ua siab ntev.
Nyem paub txoj hauv kev los txhim kho lub hlwb
Txoj kev no tsis tsuas yog ua rau kev siv lub cim xeeb ntawm LSTM rau cov ntaub ntawv keeb kwm tab sis kuj qhia txog kev cuam tshuam ntawm sab nraud ntawm tus nqi los ntawm tag nrho cov txheej txheem kev sib txuas, uas muab lub tswv yim tshiab los daws qhov teeb meem ntawm tus nqi kwv yees.Piv nrog BP neural network, lub Cov txiaj ntsig kev sim qhia tau hais tias txoj kev no muaj qhov tseeb dua thiab ruaj khov zoo dua.Tshaj tawm cov khoom lag luam piav qhia thiab cov yam ntxwv ntawm cov khoom lag luam, thiab tawm cov khoom lag luam zoo ib yam li cov khoom lag luam, ua kom tiav cov khoom lag luam cov ntaub ntawv los ntawm kev siv cov ntaub ntawv keeb kwm ntawm cov khoom zoo sib xws, thiab tsim kev cob qhia. teem los xyuas qhov tseeb ntawm txoj kev npaj.
1. Taw qhia
e kev lag luam nplai ntawm cov khoom lag luam tau nthuav dav txhua hnub, kev lag luam ntau yam tau dhau los ua nplua nuj thiab nplua nuj, thiab kev lag luam kev lag luam tau dhau los ua tus qauv ntau dua. Txij li thaum txoj kev loj hlob ntawm cov khoom lag luam, nws tau dhau los ua ib qho tseem ceeb 'nyiaj txiag submarket. eprice teeb liab kis hauv kev ua lag luam ua lub luag haujlwm tseem ceeb hauv kev coj ua lag luam tsim khoom, kev lag luam thoob ntiaj teb, thiab tswj kev lag luam [1]. Tuam Tshoj txoj kev sib pauv khoom lag luam kuj tau khaws cov ntaub ntawv lag luam nplua nuj nyob hauv ntau xyoo ntawm kev ua haujlwm thiab kev loj hlob.
Cov kws tshawb fawb sau, txheeb xyuas, thiab txheeb xyuas cov ntaub ntawv sib pauv ntawm ntau yam khoom muag thiab tom qab ntawd ntxiv thiab sau cov khoom lag luam tib yam. ese indexes muaj peev xwm Žect tag nrho lub xeev ntawm cov khoom muaj nqis Ž kev hloov pauv thiab kev loj hlob ntawm cov khoom lag luam kev lag luam thiab pab tsoomfwv cov chaw ua haujlwm nkag siab txog tus Tsov tus tw ntawm lub xeev macroeconomy; tib lub sijhawm, cov tuam txhab lag luam tseem tuaj yeem siv cov khoom lag luam tus nqi cov ntaub ntawv muaj nyob rau hauv cov khoom lag luam index. lawv tus kheej kev txiav txim siab ua lag luam, tsim nyog npaj cov khoom yuav khoom thiab txo qis kev poob nyiaj tsis tsim nyog [2, 3].
Kev ua lag luam ua lag luam nyob hauv txoj haujlwm tseem ceeb dua hauv kev lag luam hauv tebchaws. Kev coj tus cwj pwm coj mus muag tsis tsuas yog ua kom muaj txiaj ntsig tseem ceeb rau kev lag luam kev lag luam ntawm cov khoom lag luam tab sis tseem ua lub luag haujlwm tseem ceeb hauv kev sib raug zoo ntawm kev lag luam, txhim kho kev lag luam mechanisms, thiab daws cov teeb meem ntawm kev ua hauj lwm [4]. Niaj hnub no, nrog kev loj hlob ntawm cov ntaub ntawv thev naus laus zis, kev lag luam e-khoom muag, raws li tus qauv muag khoom muag tshiab, tau tsim sai heev. Cov neeg tuaj yeem tau txais ntau txhiab cov ntaub ntawv khoom muag los ntawm Is Taws Nem thiab tiv tauj cov neeg muag khoom thoob plaws ntiaj teb rau kev lag luam yam tsis tau tawm hauv lawv lub tsev [5].
Qhov tshwm sim ntawm kev lag luam e-lag luam tau ua kom yooj yim rau tib neeg lub neej, txhawb cov neeg siv khoom txaus siab, thiab nce qhov tseem ceeb ntawm cov neeg siv khoom lag luam; txoj kev loj hlob ntawm e-retail lag luam tau tsav txoj kev vam meej ntawm kev lag luam lag luam. Kaum tawm txhiab ntawm cov khw muag khoom muag tau nthuav tawm thiab muag hauv Is Taws Nem, uas txo cov nqi ntawm cov khoom muag thiab txhim kho kev muag khoom. Nyob rau tib lub sijhawm, vim tias kev lag luam e-kev lag luam cia siab rau hauv Is Taws Nem, nws muaj txiaj ntsig zoo hauv kev nrhiav tau thiab khaws cov ntaub ntawv thiab cov ntaub ntawv. Cov tuam txhab tau txais ntau cov ntaub ntawv los ntawm cov ntaub ntawv thev naus laus zis. Yuav ua li cas mine cov ntaub ntawv no thiab nrhiav cov cai muaj txiaj ntsig, coj kev txiav txim siab ntawm kev lag luam, txhim kho cov qauv muag, tsim cov tswv yim muag zoo, thiab thaum kawg tau txais txiaj ntsig kev lag luam los ntawm e-lag luam. Ntau lub tuam txhab pib nqis peev hauv kev coj ua lag luam cov ntaub ntawv tsuas thiab tsim lawv cov ntaub ntawv lag luam mining schemes.
Kev tshawb xyuas cov ntaub ntawv yog ib theem tseem ceeb ntawm kev txheeb xyuas cov ntaub ntawv, uas txawv ntawm cov ntaub ntawv thawj zaug. * Lub hom phiaj ntawm kev txheeb xyuas cov ntaub ntawv ua ntej yog nyob ntawm seb qhov yuav tsum tau ua kom pom cov qauv kev txheeb cais thiab cov lus pom zoo tau ua kom ntseeg tau tias muaj kev ntseeg siab ntawm kev txheeb xyuas qhov tseeb [6–8]. Hauv cov txheej txheem kev tshuaj ntsuam no, cov ntaub ntawv tsis tsim nyog tau sau nrog qhov tseem ceeb uas ploj lawm, hloov pauv cov ntaub ntawv, tshem tawm tus nqi qis dua, thiab lwm yam kev ua tiav txhawm rau txhim kho qhov tseeb ntawm kev tshuaj xyuas. Kev tshawb xyuas cov ntaub ntawv suav nrog kev txheeb xyuas cov ntaub ntawv ua ntej, tab sis nws qhov chaw pib tsis yog los txiav txim siab cov ntaub ntawv zoo tab sis kuj tseem yuav nrhiav pom cov qauv ntawm cov ntaub ntawv faib (Patten) thiab tawm tswv yim tshiab los ntawm cov ntaub ntawv. Kev tshawb nrhiav cov ntaub ntawv txheeb xyuas tau pom tias yog ib kauj ruam tseem ceeb hauv cov ntaub ntawv tshawb fawb ua haujlwm uas tuaj yeem cuam tshuam ntau yam txheej txheem. Nyob rau hauv cov ntaub ntawv science workflow qhia nyob rau hauv daim duab 1, exploratory dataanalysis yog ze ze rau lwm yam txheej txheem.
Txhua lub hauv paus qauv cob qhia cov tswv yim sib txawv ntawm qhov kev cob qhia cais. * Kev saib xyuas kev kawm ntxhib, txhua tus qauv. Ib qhov hnyav yog tsim los rau qhov kwv yees tus nqi ntawm kev cob qhia txheej txheem ntawm txhua tus qauv siv lub cev hnyav, thiab tom qab ntawd tus nqi tsim tawm qhov hnyav tau muab rau tus nqi kwv yees ntawm qhov ntsuas ntawm txhua tus qauv qauv [9, 10].Forecast tus nqi sib npaug thiab cuam tshuam Getmodel xeem txheej ntawm kwv yees tus nqi. Daim duab 2 qhia txog kev tsim kho thiab daim duab ntawm tus qauv. Los ntawm daim duab, peb paub tias nws feem ntau suav nrog cov ntaub ntawv ua ntej, thiab ntau qhov kev twv ua ntej tau los ntawm ntau tus qauv LSTM, thiab qhov kev twv ua ntej tau sib xyaw ua ke thiab txhim kho los ntawm qhov hnyav ntau yam tsis muaj.
Los ntawm kev txheeb xyuas tus nqi hnyav ntawm txhua lub hauv paus qauv ntawm cov khoom ua si, nws pom tau tias cov qauv hauv paus nrog qhov tseeb dua qhov tseeb yog qhov yuav tau txais qhov hnyav dua.Daim duab 3 yog txoj kev siv tshuab.
2. Cov hauj lwm cuam tshuam
Nqe kwv yees hais txog kev kwv yees tus cwj pwm ntawm dynamicanalysis ntawm yav tom ntej tus nqi hloov pauv raws li keeb kwm muaj nqis thiab tus nqi sib txawv ntawm cov khoom muag [11]. *e cov kws sau ntawv ntawm [12] tau siv cov qauv neural network rov ua haujlwm rau kev kub ceev tsheb ciav hlau kev co kwv yees los ntawm lub sijhawm series thiab ua tiav cov txiaj ntsig zoo.
Nrog rau txoj kev loj hlob ntawm lub sij hawm yooj yim series algorithm, daim ntawv thov ntawm yooj yim lub sij hawm series tsom xam yog maj expanding. Tam sim no lub sij hawm yooj yim series tsom xam algorithm tau nyob rau hauv cov nqi ua liaj ua teb, kev lag luam cov khoom lag luam priceprediction, nyiaj txiag Tshuag nqi forecasting, thiab ntau lwm yam Nws muaj ntau yam kev siv, vim yog siv tsawg nyob rau hauv kev tsom xam ntawm cov ntaub ntawv cov ntaub ntawv, tib lub sij hawm cov ntaub ntawv loweffective. tsom mus rau keeb kwm, ua rau kwv yees cov txiaj ntsig tseem tsis tuaj yeem ua tau raws li qhov xav tau ntawm kev txhim kho kev sib raug zoo.
Kev tshawb fawb cov xwm txheej ntawm tus nqi kwv yees algorithm raws li lub sijhawm yooj yim: raws li lub caij nyoog thiab lub caij nyoog ntawm lub zog tus nqi hloov pauv, Marcjasz li al. [13, 14] txheeb xyuas qhov cuam tshuam ntawm lub caij nyoog ntawm kev hloov pauv hluav taws xob yav tom ntej thiab siv NARX neural networks qauv los kwv yees tus nqi zog, uas ua tiav cov txiaj ntsig zoo. Txhawm rau kom tau txais txiaj ntsig kev kwv yees tau zoo, cov kws sau ntawv ntawm [15] tsim aDenoising Aggregation ntawm Graph neural networks los ntawm kev siv lub ntsiab lus tseem ceeb tsom xam. Wang et al. [16–18] tau kawm luv luv nqi hluav taws xob kwv yees nrog kev sib tw denoising autoencoders, ua cov kev tshawb fawb thiab kev siv lub hybrid forecasting moj khaum, thiab tau tshaj tawm cov qauv tshiab hybrid rau huab cua zoo Performance index ob-phasedecomposition txheej txheem thiab hloov kho huab kev kawm tshuab (ELM), feem. Chong et al. [19, 20] tau kawm txog kev kawm sib sib zog nqus rau kev tsom xam kev lag luam thiab kev kwv yees thiab nqa cov khoom vaj khoom tsev nqi ntawm kev kawm tshuab, feem. Nilashi et al. nthuav tawm txoj hauv kev rau kev txheeb xyuas cov ntaub ntawv loj hauv zej zog rau kev txiav txim siab ntawm cov neeg siv khoom hauv eco-friendly cov tsev so thiab tau sim cov kev daws teeb meem ntawm ob qhov qhib datasets [21]. Ntxiv rau cov haujlwm saum toj no, Hoseinzade [22] ua ke ANN qauv thiab CNN tus qauv, kwv yees qhov nce thiab poob ntawm ShanghaiFutures nyob rau lub sijhawm paub, thiab ua tiav cov txiaj ntsig zoo. Cov tub ua lag luam tseem tuaj yeem txiav txim siab nqis peev nrog kev pab los ntawm Chen thiab Ge [23] tshawb nrhiav cov txheej txheem kev xav hauv LSTM-raws li Hong Kong cov nqi twv twv txiaj. Fang et al. kev tshawb fawb txog kev nqis peev uas tsim los ntawm kev kawm tob [24]. Raws li cov lus sib tham saum toj no, cov txiaj ntsig tseem ceeb ntawm daim ntawv no tau sau tseg raws li hauv qab no:
(1) Raws li cov qauv kev txhim kho ntawm RNN qauv, tus qauv LSTM tsis tsuas yog txais cov yam ntxwv ntawm RNN qauv tsim nyog rau kev soj ntsuam nrog lub sij hawm seriesdata tab sis kuj ntxiv daws qhov teeb meem ntawm kev vam khom mus ntev rau lub sij hawm dimension thiab txhim kho qhov tseeb ntawm kev twv ua ntej. Nws cov txiaj ntsig kwv yees yog qhov zoo tshaj rau BP neural network, RNN, CNN, GRU, thiab lwm yam neural network qauv.
(2) Grid Search yog siv los cob qhia cov qauv nrog cov qauv sib txawv thiab hla-validate txhua tus qauv kom txog thaum pom qhov zoo tshaj plaws ntawm qhov tseem ceeb kom paub meej cov qauv kev ua tau zoo tshaj plaws.
(3) * e optimized qauv raug txheeb xyuas nyob rau hauv qhov kev ntsuam xyuas teeb, thiab qhov nruab nrab square yuam kev yog siv raws li qhov ntsuas ntsuas los tiv thaiv tus qauv los ntawm overfitting. * e cov txiaj ntsig qhia tau tias tus qauv ua tiav qis txhais tau tias square errorin ob qho tib si kev cob qhia thiab cov txheej txheem sim, thiab tau txais cov txiaj ntsig kev twv ua ntej.
3. LSTM Network
LSTM neural network tau xub thov los ntawm Hochreiter li al.(1997) thiab txuas ntxiv tom qab nws qhov kev ua kom zoo thiab kev txhim kho los ntawm Alex Graves. Hauv ntau qhov teeb meem uas muaj feem xyuam nrog cov ntaub ntawv sib lawv liag, LSTM tau ua tiav zoo thiab tau siv dav, xws li kev ua cov lus ntuj (NLP), kev twv ua ntej lub sijhawm, thiab lwm yam.
Cov tsoos recurrent neural network (RNN) tsis tuaj yeem daws qhov teeb meem mus sij hawm ntev, los daws qhov teeb meem no, LSTM neural network tau thov los ntawm kev ntxiv "lub rooj vag" qauv los tswj lub xeev ntawm tes thiab cov khoom tawm ntawm lub sijhawm sib txawv los daws qhov teeb meem ntawm gradient ploj. Tus qauv "gate" ntawm LSTM suav nrog peb hom: "tsis nco qab lub rooj vag," "qhov rooj nkag," thiab "qhov rooj tawm." Kev ua haujlwm ntawm "tsis nco qab lub rooj vag" yog los txiav txim cov ntaub ntawv xa tawm los ntawm lub sijhawm dhau los mus rau lub sijhawm tam sim no thiab xaiv "tsis nco qab" qee cov ntaub ntawv raws li qhia hauv daim duab 4. Ntxiv rau, qhov pib orthogonal tau npaj tseg kom tsis txhob gradient ploj los yog tawg ntawm thawj theem ntawm kev cob qhia. , ReLU (Recti'ed Linear Unit) ua kom muaj nuj nqi tuaj yeem txo qis kev ploj mus, gradient shear cansolve gradient explosion, thiab LSTM Unit tuaj yeem tswj kev ploj mus.
LSTM tau ua tiav tiav hauv tshuab txhais lus, kev sib tham tiam, thiab lwm yam'fields, uas qhia txog kev ua qauv zoo heev ntawm cov ntaub ntawv sib lawv liag. Yog li, daim ntawv no tsim tus nqi muag khoom ntawm tus qauv kev ntaus pob ncaws pob raws li LSTM network chav tsev thiab tuaj yeem siv tag nrho nws cov yam ntxwv uas txhua qhov ntev tuaj yeem siv tau los ua cov tswv yim thiab siv rau kev paub txog cov ntaub ntawv online. LSTM daws qhov teeb meem uas RNN tsis tuaj yeem ua haujlwm ntev ntev los ntawm kev qhia [25].
Cia tus naj npawb ntawm cov input neurons nyob rau hauv tag nrho cov zais zais G, G suav nrog tag nrho cov chav nyob thiab lub rooj vag, thiab siv qhov Performance index G torepresent cov input neurons. e rau pem hauv ntej xam ntawm LSTM yog xam ib tug input sequence X nrog ib tug ntev ntawm lub sij hawm d, nws pib taw tes yog t 1 [26, 27]. Thaum tus nqi ntawm lub sij hawm taw tes T nce tsis tu ncua, qhov sib npaug yuav recursively mus txog t. Zoo li kev suav rau pem hauv ntej, rov qab suav yog ib qho kev sib txuas lus X nrog lub sijhawm ntev ntawm T, tab sis qhov pib ntawm kev rov qab suav yog T T.Thaum tus nqi ntawm T txo qis tsis tu ncua, lub reciprocalof chav tsev yog xam recursively mus txog 1 1. Raws li cov derivatives ntawm txhua lub sij hawm taw tes saum toj no, peb tuaj yeem tau txais tus nqi 'nal nyhav derivative.
qhov twg Wf thiab bf sawv cev qhov hnyav thiab kev tsis ncaj ncees ntawm lub rooj vag tsis nco qab, feem, thaum σ sawv cev rau Sigmoidfunction.
4. Kev ua cov ntaub ntawv thiab tshawb xyuas
ntawm * Lub ntsiab lus tshiab ntawm daim ntawv no yog txhawm rau txhim kho tus qauv LSTM cov qauv thiab siv nws los kwv yees tus nqi muag khoom ntawm cov khoom kis las, yog li peb feem ntau muab piv nrog cov qauv LSTM. Thib ob, vim tias cov ntaub ntawv tiag tiag muaj nuj nqis thiab nyuaj kom tau txais, daim ntawv no tsuas yog siv cov ntaub ntawv nkaus xwb rau qhov simulation simulation. Vim tias qhov zoo ntawm cov ntaub ntawv cuam tshuam rau kev cob qhia ntawm cov qauv xaiv, kev sau, kev txheeb xyuas, thiab kev ua cov ntaub ntawv yog cov theem tseem ceeb ua ntej kev cob qhia qauv. *e cov ntaub ntawv hauv daim ntawv no feem ntau suav nrog ob ntu: cov khoom tshawb fawb thiab cov ntaub ntawv yam ntxwv.
Kev kwv yees dav dav ntawm cov khoom lag luam tus nqi tsom rau cov khoom siv kis las. Siv Python los ua qhov piav qhia txog kev txheeb xyuas cov khoom sib txawv ntawm cov khoom ua si thiab kos lawv cov kab kos nqe kaw [31, 32], raws li pom hauv daim duab 5. * e-kev faib tawm yog skewed rau sab xis, lub ncov me dua 3, tus Tsov tus tw yog nyias, thiab tsis ua raws li qhov kev faib tawm.
n.4.1. Cov ntaub ntawv suab nrov txo. Raws li kev lag luam kev lag luam yog qhov nyuaj heev, cov ntaub ntawv no muaj cov suab nrov tsis tu ncua, yog li lub tsev qiv ntawv hauv Python yog siv rau kev hloov pauv wavelet kom tshem tawm cov ntaub ntawv nrov [33]. Nws yog ib nqi sau cia hais tias cov pa waveletchange qauv yog siv nyob rau hauv daim ntawv no. Nws inherits thiab txhim kho lub tswv yim ntawm lub sij hawm luv luv Fourier hloov localization thiab overcomes lub shortcomings ntawm qhov rais loj tsis hloov nrog zaus. Nws tuaj yeem muab qhov "sijhawm-zaum" qhov hloov pauv nrog zaus, uas yog ib qho cuab yeej zoo tshaj plaws rau kev soj ntsuam lub sij hawm-zaus thiab ua cov cim. * Yog li ntawd, nws yog qhov tshwj xeeb tshaj yog tsim nyog rau kev tshem tawm suab nrov hauv cov ntaub ntawv nyiaj txiag.Daim duab 6 thiab 7 yog qhov sib piv ua ntej thiab tom qab wavelettransform.
n.Tom qab kev cob qhia LSTM neural network tiav lawm, cov txiaj ntsig kev twv ua ntej thiab cov txiaj ntsig MSE sib raug zoo ntawm Model 1 thiab Model 2 tau muab, raws li qhia hauv daim duab 8, thiab cov txiaj ntsig kev twv ua ntej ntawm Model 1 thiab Model 2 muaj nyob hauv daim duab 9.
9.* Lub tswv yim tseem ceeb ntawm cov qauv kev cob qhia yog kom haum rau cov txheej txheem kev cai ntawm cov ntaub ntawv qhia kev cob qhia kom nthuav tawm cov cai hauv cov ntaub ntawv.Ib lo lus, qhov haum piav qhia tias zoo npaum li cas, lossis zoo npaum li cas, themodel tuaj yeem nthuav dav rau cov ntaub ntawv. nyob rau hauv lub xeem teeb Ib tus qauv zoo ua rau cov qauv kev ua tau zoo thiab tuaj yeem siv tau nrog cov ntaub ntawv tshiab sab nraum cov ntaub ntawv qhia kev cob qhia, piv txwv li, cov ntaub ntawv tawm ntawm cov qauv. Ntxiv rau qhov tsis muaj peev xwm ua tau zoo, cov qauv sib txawv yuav tsum muaj qhov sib txawv hyperparameters, uas yog qhov tsis tas yuav tsum tau kawm.
Parameters yog qhov tseem ceeb rau tus qauv thiab nyob ntawm qhov kev cob qhia cov ntaub ntawv. Raws li ib feem ntawm cov txheej txheem kev cob qhia, tus qauv LSTM tau hloov kho thiab ua kom zoo dua kom tau txais kev kwv yees zoo dua los ntawm kev kawm tsis tau los ntawm kev cob qhia cov ntaub ntawv los ntawm cov txheej txheem kev ua kom zoo [34–37].
Cov txheej txheem ntawm kev tshawb fawb txog kev lag luam feem ntau yog piav qhia txog kev lag luam phenomena los ntawm kev tsim cov qauv kev lag luam.Thaum feem ntau txais kev lag luam phenomena yog con-'rmed los ntawm cov kws tshawb fawb, lawv yuav tsum tau teev nyob rau hauv daim ntawv ntawm cov lus tseeb. Hauv kev tshawb nrhiav tom qab, yog tias qhov kev daws teeb meem ntawm themodel hauv lub xeev kev sib npaug sib luag yog raws li qhov tseeb, nws tuaj yeem piav qhia tias tus qauv tsim nyog ntau dua rau qhov zoo. Nyob rau hauv txoj kev tshawb no ntawm 'cov teeb meem nyiaj txiag, tshwj xeeb tshaj yog nyob rau hauv txoj kev tshawb no ntawm 'nyiaj txiag lub sij hawm rov qab series, ib co kev txheeb cais cov yam ntxwv muaj peev xwm feem ntau raug soj ntsuam.Raws li cov kws tshawb fawb' raug cov lus tseeb ntawm univariate returnseries, lawv yog summarized raws li nram no: lawv feem ntau showautocorrelation; Nws feem ntau qhia txog kev nco ntev; e qeeb decayof absolute rov autocorrelation; ncov tuab tail tis; qhov kev faib tawm hloov pauv raws sijhawm; waveaggregation e‡ect; Tom qab kho cov Žactuation aggregation, tseem muaj ib tug conditional tuab tail e‡, thiab lwm yam. Nyob rau hauv lub assumption ntawm ywj siab thiab idenically faib, qhov kev ua tau zoo ntawm lub 'kuaj' nyiaj txiag lub sij hawm series qauv yog feem ntau tsis pom zoo, yog li ntawd tus yam ntxwv faib ntawm cov dataneed yuav tsum tau txiav txim siab. thaum ua qauv ntawm cov nyiaj tau los series.Thaum tus qauv neural network siv los ua qauv cov ntaub ntawv, qhov kev xav ntawm kev faib khoom tsis tas yuav tsum tau txiav txim siab. yog vim hais tias tus qauv neural network tuaj yeem nthuav dav cov qauv ntawm cov ntaub ntawv tawm tswv yim kom cov yam ntxwv ntawm 'cov ntaub ntawv nyiaj txiag tuaj yeem ntes los ntawm neural network.
Grid Search yog ib txoj hauv kev ntawm kev cob qhia amodel, siv cov kev sib txawv ntawm hyperparametervalues los cob qhia tus qauv, hla kev lees paub txhua tus qauv kom txog rau thaum pom kev sib xyaw ua ke ntawm qhov tseem ceeb los xyuas kom meej qhov kev ua tau zoo tshaj plaws. Los ntawm kev ntsuas tsis tu ncua ntawm txhua qhov kev sib txuas, ib pab pawg rau 'thiab qhov tsim nyog tshaj plaws ntawm kev sib txuas ntawm super configuration parameters yuav discretization, super tsis raws li lawv cov yam ntxwv los xaiv ntau qhov kev paub dhau los, thiab tom qab ntawd raws li kev sib txawv ntawm cov qauv kev cob qhia, xaiv qhov kev pom zoo ua ke ntawm con'guration, qhov xwm txheej ntawm tsawg haum rau super tsis. Random Search yog randomly combinehyperparameters thiab ces xaiv qhov pom configuration.Nws tsis ua rau tsis tsim nyog sim ntawm unimportantparameters vim hais tias, ib yam li cov tsis tu ncua coefficients muaj alimited cuam tshuam rau cov qauv kev ua tau zoo, cov kev kawm cov nqi muaj ntau dua cuam tshuam rau cov qauv kev ua tau zoo, yog li nws tsis tsim nyog sim. Kev tshawb nrhiav random feem ntau muaj txiaj ntsig thiab siv tau yooj yim dua li kev tshawb nrhiav kab sib chaws.Txawm li cas los xij, ob txoj hauv kev no tsis xav txog seb puas muaj kev sib raug zoo ntawm hyperparameters, yog li lawv kuj tsis muaj txiaj ntsig. Bayesian optimization yog ib qho kev hloov kho hyperparameter optimization txoj kev, uas kwv yees tias tom qab ntawd ua tau hyperparameter ua ke raws li kev kuaj pom hyperparameter ua ke kom tau txais txiaj ntsig siab tshaj plaws. Txij li thaum qhov kev faib ua feem ntawm Gaussian faib yog s-hom muaj nuj nqi, GELUfunction tuaj yeem kwv yees los ntawm tanh muaj nuj nqi lossis Logisticfunction raws li qhia hauv daim duab 10.
Lub tswv yim yooj yim tshaj plaws yog 'xa kawm tus nqi thoob plaws hauv txoj kev cob qhia. Xaiv qhov kev kawm me me tso cai rau lub optimizer 'nd ib qho kev daws teeb meem zoo, tab sis nws yog ib qho yooj yim los txwv qhov sib koom ua ke. Kev sib raug zoo ntawm ob leeg tuaj yeem sib npaug los ntawm kev siv sij hawm los hloov qhov kev kawm.Daim duab 11 qhia qhov kev kawm ntawm txhua lub sijhawm.
Tom qab ntau zaus ntawm kev hloov kho thiab kev ua kom zoo, qhov kev twv ua ntej thiab tom qab kev txhim kho tau pom nyob rau hauv daim duab 12 thiab 13, raws li. e 'nal qauv qauv thiab cov tsis tau txais yog raws li nram no: qhov Sequentiallength ntawm qhov ntsuas qhov rais yog 55, thiab cov sequential qauv muaj peb LSTM khaubncaws sab nraud povtseg, nrog rau cov xov tooj ntawm neurons ineach txheej yog 100, 100, thiab 150, raws. Toavoid over-fitting, ob Dropout txheej yog ntxiv nrog lub Dropout txheej ntawm 0.2, thiab qhov ntev ntawm cov ntaub ntawv tawm tswv yim yog 5. e tuab txheej tau ntxiv rau aggregate nws dimension rau hauv 1, kev ua kom muaj nuj nqi yog linear, thiab qhov poob muaj nuj nqi wasset as Mean Squared yuam kev (MSE). Adas tau siv los ua qhov optimization algorithm, thiab ob Epochs tau siv los ua tus qauv. Txhua batch yog 32 loj.
e cov qauv piv txwv tus nqi kwv yees kuj yog tus qauv kev kawm tob. Ua ntej, tus qauv kev twv ua ntej tus nqi yog tsim; uas yog, tus qauv tsuas muaj marketdata thiab 'cov ntaub ntawv nyiaj txiag. e yav dhau los cov ntaub ntawv kev lag luam thiab '-cov ntaub ntawv nyiaj txiag yog siv los cob qhia cov qauv kev kawm sib sib zog nqus kom tau txais cov qauv kev kawm sib sib zog nqus los kwv yees lub lag luam yav tom ntej; rough natural language processing thev naus laus zis, kev tshem tawm cov ntaub ntawv kev xav, thiab kev ntsuam xyuas kev xav tau ua los ntawm cov ntaub ntawv pej xeem kev xav. Ua ke nrog cov qauv kev kwv yees tus nqi sib piv, cov qauv kev kawm tob tob uas siv cov ntaub ntawv kev lag luam, 'cov ntaub ntawv nyiaj txiag, cov ntaub ntawv tshawb fawb, thiab cov ntaub ntawv xav tau ntawm ' xov xwm nyiaj txiag, thiab cov qauv siv los kwv yees lub lag luam yav tom ntej. en, qhov kev twv ua ntej e‡ect ntawm cov qauv kev nyab xeeb tus nqi kwv yees piv nrog cov qauv kev kawm sib sib zog nqus raws li cov lus ua tau zoo ntawm cov ntaub ntawv pej xeem kev xav.
Tom qab txiav txim tus naj npawb ntawm cov input nodes, outputnodes, thiab zais txheej ntawm ob tus qauv LSTMnetwork, cov qauv kev kawm tob tuaj yeem raug cob qhia. Tom qab ntau qhov kev sim, nws pom tias lub sijhawm kev cob qhia tsawg dhau lawm thiab cov qauv kev cob qhia yuam kev yog qhov loj heev, yog li nws yog qhov tsim nyog yuav tsum tau ua kom cov sij hawm cob qhia tsis tu ncua, tab sis nrog kev nce ntawm cov sij hawm kawm; qhov yuam kev ntawm cov qauv kev cob qhia maj mam nyhav rau tus nqi ruaj khov. Yog hais tias lub sij hawm kev cob qhia ntawm themodel yog nce ntau heev nyob rau lub sij hawm no, tus qauv e‡, thiab lwm yam yog notimproved ntau. Nws yog qhov tshwm sim ntawm kev cob qhia 'tus qauv so ntawm C. Thaum lub sij hawm kev cob qhia tsawg dua 200, qhov yuam kev ntawm tus qauv yog loj. Nyob rau lub sijhawm no, nce lub sijhawm kev cob qhia yuav txo qis qhov yuam kev ntawm kev cob qhia qauv; Thaum lub sij hawm kev cob qhia ntau tshaj 200 zaug thiab tsawg dua 1000 zaug, tus qauv yuam kev tau me me. Lub sijhawm no, thaum lub sijhawm kev cob qhia tau nce, qhov txo qis ntawm cov qauv kev cob qhia yuam kev tau qhia txog kev txo qis; Thaum kev cob qhia 1000 txog 2000 zaug, qhov yuam kev ntawm tus qauv hloov pauv hauv thaj chaw me me, thiab cov txiaj ntsig ntawm kev nce cov sijhawm ntawm kev cob qhia yog maj mam tsis pom tseeb. rough manytests and comparisons, nws pom tau hais tias thaum tus naj npawb ntawm kev cob qhia lub sij hawm ntawm tus qauv yog hais txog 2000 lub sij hawm, nws muaj peev xwm ua tau raws li qhov tseeb cov kev cai ntawm kev cob qhia. Yog hais tias tus naj npawb oftraining lub sij hawm yog nce, lub e‡ect ntawm kev txhim kho cov modeltraining yuam kev yog me me, Ntxiv mus, nws yuav siv sij hawm ntev los cob qhia tus qauv ntawm lub computer, yog li nws yog ib qho tseem ceeb me ntsis kom nce ntau lub sij hawm kawm. Yog li ntawd, peb teem lub sijhawm kawm ntawm cov qauv kev kawm tob hauv qhov kev kawm no ntawm 2000 zaug. Hauv cov qauv tom qab, peb kuj txheeb xyuas tias nws ua raws li txoj cai no. Yog li ntawd, tus qauv kev kawm qhov tob thib ob nrog rau cov ntaub ntawv qhia txog pej xeem cov tswv yim kuj tau teem sijhawm rau kev cob qhia ntawm 2000 zaug.
5. Cov lus xaus
(1) Daim ntawv no siv cov kev tshawb xav tob tob, raws li cov yam ntxwv ntawm 'cov ntaub ntawv nyiaj txiag lub sij hawm series, siv LSTM neural network qauv los kwv yees cov khoom lag luam tus nqi ntsuas, thiab sib piv nws cov txiaj ntsig kwv yees nrog cov txiaj ntsig kwv yees siv lub network qauv. Cov txiaj ntsig kev sim qhia tau tias tus qauv LSTM neuralnetwork muaj qhov ua tau zoo tshaj plaws ntawm qhov ntsuas.
(2) Cov qauv kev twv ua ntej raug tsim. Raws li collated cov ntaub ntawv feature engineering, ib tug neural networkprediction qauv raws li ntev thiab luv-termmemory yog tsim, thiab cov qauv raug cob qhia nrog ib tug kev cob qhia teem rau kwv yees tus nqi ntawm cov khoom kis las. Interms ntawm tus qauv optimization, tus naj npawb ntawm cov hindlayer neurons, kev kawm tus nqi, batch loj, thiab kev cob qhiawheel raug kho kom ua tiav qhov zoo tshaj plaws kev cob qhia.
(3) Nyob rau hauv cov khoom kis las no tus nqi sib tw twv qauv, cov ntaub ntawv sib txawv ntawm tus nqi kis las yuav muaj kev cuam tshuam sib txawv ntawm qhov kev twv ua ntej, yog li kev xaiv ntawm cov ntaub ntawv teev kuj tseem ceeb heev.
Txawm hais tias tus qauv npaj rau hauv daim ntawv no ua tiav cov txiaj ntsig kwv yees tau zoo, tus qauv tsis xav txog qhov sib cuam tshuam ntawm cov ntaub ntawv sijhawm. Qee lub sij hawm zawv zawg qhov rais cov cuab yeej siv tau rau hauv cov kev tshawb fawb yav tom ntej los txhim kho cov kauj ruam kwv yees loj thiab qhov tseeb ntawm tus qauv.
Cov ntaub ntawv muaj
Cov ntaub ntawv siv los txhawb qhov kev tshawb pom ntawm qhov kev tshawb fawb no muaj los ntawm tus sau thaum thov.
Kev tsis sib haum xeeb
Covtus sau tshaj tawm tias tsis muaj kev tsis sib haum xeeb ntawm kev txaus siab.
Cov ntaub ntawv
[1] J. Wang, R. Hou, thiab C. Wang, "Txhim kho kev txhawb nqa vectorregression qauv raws li kev xaiv sib txawv thiab brainstormoptimization rau Tshuag nqi kwv yees," Applied Soft Computing, vol. 4 9, nr. 2, pp. 164–178, 2016.
[2] M. Dixon, D. Klabjan, thiab J. Hoon Bang, "C1assificationbased kev lag luam nyiaj txiag kwv yees siv cov sib sib zog nqus neural," Algorithmic Finance, vol. 6, nre. 3-4, p. 67–77, 2017.
[3] K. Chen, Y. Zhou, thiab FY Dai, "Ib qho LSTM-raws li txoj kev forstock rov los kwv yees: cov ntaub ntawv kawm ntawm Tuam Tshoj Tshuag lag luam,"in Proceedings of the 2015 IEEE International Conference onBig Data, vol. 123, Nr. 12, p. 2823-2824, Santa Clara, CA, USA, 2015.
[4] AE Clements thiab N. Toodorova, "Information flow, tradingactivity and commodity futures volatility," Journal of FuturesMarkets, vol. 3 6, nr. 1, pp. 88–104, 2016.
[5] L. Xuewei thiab L. Xueyan, "Cov ntaub ntawv loj thiab nws cov thev naus laus zis tseem ceeb yav tom ntej," Kev suav hauv Science & Engineering, vol. 20, nr. 4 ,pab. 75–88, 2018.
[6] AI Sarwat, M. Amini, A. Domijan, A. Damnjanovic, F. Kaleem, "Kev kwv yees huab cua-raws li kev cuam tshuam hauv cov phiaj xwm thev naus laus zis siv cov ntaub ntawv tshawb fawb," Journal of ModernPower Systems thiab Clean Energy, vol. 4, nre. 2, pp. 308–315, 2016.
[7] X. Zhang, J. Wang, thiab Y. Gao, "Ib tug hybrid luv luv nqi hluav taws xob forecasting moj khaum: cuckoo search-basedfeature xaiv nrog singular spectrum tsom xam thiab SVM,"Energy Economics, vol. 8 1, nr. 6, pp. 16–23, 2019.
[8] G. Osorio, M. Lotfi, VMA Campos, thiab M. Khan, "Hybrid 'forecasting qauv rau luv luv nqi hluav taws xob ua lag luam nrog rau kev sib txuas ntxiv," Sustainability, vol. 11, nr. 1 ,pab. 1-15, 2018.
[9] A. Bello, DW Bunn, thiab J. Reneses, "Kev kwv yees nruab nrab-Termprobabilistic ntawm cov nqi hluav taws xob: ib qho kev sib txuas," IEEE Transactions on Power Systems, vol. 16, nr. 2 ,pab. 1–6, 2016.
[10] A. Bello, J. Reneses, A. Muñoz, thiab A. Delgadillo, "Probabilistic forecasting of hourly hluav taws xob nqi nyob rau hauv lub sij hawm nruab nrab siv spatial interpolation cov tswv yim," InternationalJournal of Forecasting, vol. 3 2, nr. 3, pp. 966–980, 2016.
[11] P. Kou, D. Liang, L. Gao, thiab J. Lou, "Probabilistic electronicprice forecasting with variational heteroscedastic Gaussianprocess and active learning," Energy Conversion and Management, vol. 89, pp. 298–308, 2015.
For more information:1950477648nn@gmail.com
