国产日韩欧美一区二区三区三州_亚洲少妇熟女av_久久久久亚洲av国产精品_波多野结衣网站一区二区_亚洲欧美色片在线91_国产亚洲精品精品国产优播av_日本一区二区三区波多野结衣 _久久国产av不卡

?

基于WOA-LSTM的窄帶通信網(wǎng)網(wǎng)絡(luò)時延預(yù)測算法

2022-02-14 05:13蘇鵬飛徐松毅于曉磊
河北工業(yè)科技 2022年1期

蘇鵬飛 徐松毅 于曉磊

摘 要:為了給窄帶通信網(wǎng)的鏈路選擇及協(xié)議的智能切換提供實時參考,設(shè)計了一種基于鯨魚優(yōu)化算法(WOA)和長短期記憶神經(jīng)網(wǎng)絡(luò)(LSTM)的窄帶通信網(wǎng)網(wǎng)絡(luò)時延預(yù)測算法。首先對實測數(shù)據(jù)樣本進行標準化處理,以LSTM神經(jīng)網(wǎng)絡(luò)算法的均方根誤差函數(shù)的倒數(shù)作為適應(yīng)度函數(shù);其次采用鯨魚優(yōu)化算法對LSTM神經(jīng)網(wǎng)絡(luò)的學(xué)習(xí)率、隱含層神經(jīng)元個數(shù)進行優(yōu)化,最后將全局最優(yōu)解輸出作為LSTM神經(jīng)網(wǎng)絡(luò)的初始參數(shù)對樣本進行訓(xùn)練預(yù)測。結(jié)果表明,基于WOA-LSTM的網(wǎng)絡(luò)時延預(yù)測算法預(yù)測精度相較于LSTM神經(jīng)網(wǎng)絡(luò)算法和BP神經(jīng)網(wǎng)絡(luò)算法分別提高了14.87%和78.89%,WOA-LSTM達到收斂時迭代次數(shù)相較于LSTM神經(jīng)網(wǎng)絡(luò)算法減少了11.11%。所提算法新穎可靠,可更準確地進行網(wǎng)絡(luò)時延預(yù)測,為窄帶通信網(wǎng)網(wǎng)絡(luò)的智能化與自動化升級提供數(shù)據(jù)支持。

關(guān)鍵詞:計算機神經(jīng)網(wǎng)絡(luò);鯨魚優(yōu)化算法;LSTM神經(jīng)網(wǎng)絡(luò);窄帶通信網(wǎng);網(wǎng)絡(luò)時延預(yù)測

中圖分類號:TN915.1 ? 文獻標識碼:A ? DOI: 10.7535/hbgykj.2022yx01002

Abstract:In order to provide real-time reference for link selection and protocol intelligent switching in narrowband communication networks,a network delay prediction algorithm based on whale optimization algorithm (WOA) and long short-term memory (LSTM) was designed.Firstly,the measured data samples were standardized,and the reciprocal of root mean square error function of LSTM neural network algorithm was used as fitness function.Secondly,the whale optimization algorithm was used to optimize the learning rate and the number of hidden layer neurons of LSTM neural network.Finally,the output of global optimal solution was used as the initial parameter of LSTM neural network to train and predict samples.The results show that compared with LSTM neural network algorithm and BP neural network algorithm,the prediction accuracies of network delay prediction algorithm based on WOA-LSTM are improved by 14.87% and 78.89% respectively,and the iteration times of WOA-LSTM are reduced by 11.11% compared with LSTM neural network algorithm when WOA-LSTM reaches convergence.The algorithm is novel and reliable,which can predict network delay more accurately and provide data support for intelligent and automatic upgrade of narrowband communication networks.

Keywords:computer neural network;whale optimization algorithm;LSTM neural network;narrowband communication network;network delay prediction

窄帶通信網(wǎng)絡(luò)是為某些特殊場景提供應(yīng)急通信保障的低速通信系統(tǒng)的主要構(gòu)成部分,其網(wǎng)絡(luò)時延受到網(wǎng)絡(luò)拓撲結(jié)構(gòu)、氣象變化因素、網(wǎng)絡(luò)協(xié)議及路由算法等多方面因素影響,當網(wǎng)絡(luò)拓撲結(jié)構(gòu)、網(wǎng)絡(luò)協(xié)議及路由算法固定下來之后,時間序列成為誘導(dǎo)其變化的主要影響因子。傳統(tǒng)的窄帶通信網(wǎng)網(wǎng)絡(luò)協(xié)議單一,根據(jù)需求需要手動進行鏈路選擇,隨著窄帶通信網(wǎng)的網(wǎng)絡(luò)復(fù)雜度增加及多種網(wǎng)絡(luò)協(xié)議的接入,迫切需要通過對窄帶通信網(wǎng)網(wǎng)絡(luò)時延預(yù)測,從而為窄帶通信網(wǎng)的鏈路選擇及網(wǎng)絡(luò)協(xié)議的切換提供參考。目前,網(wǎng)絡(luò)時延預(yù)測主要有基于數(shù)理統(tǒng)計的數(shù)學(xué)建模法,最小二乘支持向量機,神經(jīng)網(wǎng)絡(luò)算法。文獻[1]通過對統(tǒng)計數(shù)據(jù)的回歸分析和誤差分析,提出了一種基于自回歸求和滑動平均(ARIMA)模型,對網(wǎng)絡(luò)化控制系統(tǒng)的隨機時延進行預(yù)測,相較于ARMA模型精度有所提高;文獻[2]提出了一種基于粒子群算法優(yōu)化(PSO)的最小二乘法支持向量機(LS-SVM)算法,對列車通信網(wǎng)絡(luò)的網(wǎng)絡(luò)時延進行預(yù)測,但是PSO優(yōu)化的參數(shù)維度較高,會影響預(yù)測時效性;文獻[3]運用BP神經(jīng)網(wǎng)絡(luò),同時運用PSO算法對神經(jīng)網(wǎng)絡(luò)權(quán)值和閾值進行優(yōu)化,通過機器學(xué)習(xí)的方法對歸一化的網(wǎng)絡(luò)時延數(shù)據(jù)進行預(yù)測,但BP神經(jīng)網(wǎng)絡(luò)沒有記憶性的特點,使得其只能通過前兩個時序的時延數(shù)據(jù)預(yù)測下一時刻的網(wǎng)絡(luò)時延,無法關(guān)聯(lián)前面更長時間時序數(shù)據(jù)的特征。對此,本文選取單一對流層散射通信鏈路構(gòu)成的窄帶通信網(wǎng)絡(luò),提出了長短期記憶神經(jīng)網(wǎng)絡(luò)(LSTM)算法,關(guān)聯(lián)長短期各個時序的網(wǎng)絡(luò)時延的歷史數(shù)據(jù),通過鯨魚優(yōu)化算法(WOA)優(yōu)化LSTM神經(jīng)網(wǎng)絡(luò)的學(xué)習(xí)率,隱含層神經(jīng)元個數(shù)和最大訓(xùn)練次數(shù),提高算法預(yù)測精度,對其網(wǎng)絡(luò)時延進行預(yù)測。

1 數(shù)據(jù)采集和預(yù)處理

1.1 數(shù)據(jù)采集

通過野外試驗,搭建了對流層散射通信鏈路組成的通信網(wǎng)絡(luò),每隔30 min在收發(fā)兩端進行大小為64 B的數(shù)據(jù)包傳輸測試,共記錄了由300組網(wǎng)絡(luò)時延數(shù)據(jù)所組成的一維實驗數(shù)據(jù)。

1.2 數(shù)據(jù)預(yù)處理

將采集到的數(shù)據(jù)進行標準化處理,將數(shù)據(jù)處理成均值為0,標準差為1的標準化數(shù)據(jù)。在神經(jīng)網(wǎng)絡(luò)的反向傳播過程中,采用了梯度下降法更新權(quán)值以及偏置值,將數(shù)據(jù)進行標準化處理可以提升模型的收斂速度,也避免了數(shù)值輸入過大,導(dǎo)致更新過程中梯度過大從而使網(wǎng)絡(luò)的學(xué)習(xí)停止更新,設(shè)置學(xué)習(xí)率時也不必再根據(jù)輸入值的范圍進行調(diào)整。樣本數(shù)據(jù)的標準化處理公式如下:

2.3 WOA-LSTM網(wǎng)絡(luò)時延預(yù)測模型

LSTM神經(jīng)網(wǎng)絡(luò)的網(wǎng)絡(luò)訓(xùn)練效果以及訓(xùn)練過程中的擬合速度和初始的參數(shù)設(shè)置密切相關(guān),其中學(xué)習(xí)率和神經(jīng)網(wǎng)絡(luò)隱含層神經(jīng)元個數(shù)直接影響了神經(jīng)網(wǎng)絡(luò)的訓(xùn)練精度和收斂速度[11-13]。對于學(xué)習(xí)率的設(shè)置來說,若初始學(xué)習(xí)率設(shè)置的過大,會導(dǎo)致偏離值較大且到后期無法擬合,學(xué)習(xí)率設(shè)置的過小,收斂速度會很慢[15]。對于隱含層節(jié)點個數(shù)來說,設(shè)置過少會欠擬合,過多會導(dǎo)致過擬合[16]。通過鯨魚優(yōu)化算法,全局向局部搜索尋優(yōu),確定最優(yōu)學(xué)習(xí)率和隱含層神經(jīng)元個數(shù),從而進行神經(jīng)網(wǎng)絡(luò)的訓(xùn)練。WOA-LSTM的網(wǎng)絡(luò)時延預(yù)測模型流程如圖4所示。

通過鯨魚優(yōu)化算法(WOA)來優(yōu)化長短期記憶神經(jīng)網(wǎng)絡(luò)(LSTM),只通過LSTM神經(jīng)網(wǎng)絡(luò)進行網(wǎng)絡(luò)訓(xùn)練前,需要由經(jīng)驗手動設(shè)置神經(jīng)網(wǎng)絡(luò)的學(xué)習(xí)率和神經(jīng)網(wǎng)絡(luò)隱含層神經(jīng)元個數(shù),通過不斷的嘗試,得到可使神經(jīng)網(wǎng)絡(luò)預(yù)測精度相對較高的參數(shù)搭配,但是很難得到在一定范圍內(nèi)使得神經(jīng)網(wǎng)絡(luò)預(yù)測精度最高的最佳參數(shù)設(shè)置;引入鯨魚優(yōu)化算法,首先設(shè)置兩種參數(shù)的搜索范圍,然后經(jīng)過上述描述的鯨魚優(yōu)化算法在此范圍內(nèi)進行隨機搜索,得到的預(yù)測誤差即損失函數(shù)TrainingLoss不斷收斂,達到精度要求時,得出最優(yōu)參數(shù),進而完成LSTM神經(jīng)網(wǎng)絡(luò)的參數(shù)初始化。

本文采用平均絕對百分比誤差(MAPE)作為鯨魚算法的損失函數(shù)。當損失值達到事先設(shè)置的下限時,得到優(yōu)化參數(shù)值。損失函數(shù)TrainingLoss的定義式如下:

TrainLoss=MAPE(h,y)=1n∑ni=1|h(i)-y(i)y(i)|,

式中:h(i)是預(yù)測結(jié)果中的第i個預(yù)測值;y(i)是數(shù)據(jù)樣本中第i個真實值;n為預(yù)測樣本數(shù)。預(yù)測值越是精確,得到的損失值越小。

預(yù)測流程如下。

步驟1:以1.1節(jié)和1.2節(jié)所述方法對數(shù)據(jù)進行處理,并以前一時間步數(shù)據(jù)預(yù)測后一時間步的數(shù)據(jù)格式輸入到WOA-LSTM模型中;

步驟2:初始化LSTM模型參數(shù)學(xué)習(xí)率和神經(jīng)網(wǎng)絡(luò)隱含層神經(jīng)元個數(shù);

步驟3:鯨魚算法種群初始化。將(n,ε)這兩個變量組成的一組值作為待優(yōu)化參數(shù)輸入到鯨魚算法中,n代表神經(jīng)網(wǎng)絡(luò)隱含層神經(jīng)元個數(shù),ε代表學(xué)習(xí)率;

步驟4:將初始化后的值作為歷史最優(yōu)值對LSTM的參數(shù)賦值并訓(xùn)練;

步驟5:將使用傳統(tǒng)LSTM訓(xùn)練得到的TrainingLoss設(shè)置為系統(tǒng)要求的終止值,并求取經(jīng)過鯨魚算法優(yōu)化后的模型損失值;

步驟6:若經(jīng)過鯨魚算法優(yōu)化后的模型損失值小于TrainingLoss,則滿足要求,利用訓(xùn)練好的模型迭代輸出網(wǎng)絡(luò)時延預(yù)測值;

步驟7:若損失值無法小于TrainingLoss或者迭代次數(shù)未到最大,則更新參數(shù)并且重新進行訓(xùn)練。

3 仿真分析

3.1 實驗設(shè)置

為了充分驗證所提出的WOA-LSTM模型在網(wǎng)絡(luò)時延預(yù)測上的有效性,設(shè)計了WOA-LSTM和LSTM神經(jīng)網(wǎng)絡(luò)以及BP神經(jīng)網(wǎng)絡(luò)預(yù)測的對比實驗,通過鯨魚優(yōu)化算法來優(yōu)化LSTM模型的最佳學(xué)習(xí)率和隱藏層單元數(shù),WOA算法在迭代過程中不斷地調(diào)整初始化LSTM模型參數(shù),直到調(diào)整到誤差值較小的LSTM神經(jīng)網(wǎng)絡(luò)模型。同時,引入BP神經(jīng)網(wǎng)絡(luò)對時延數(shù)據(jù)進行訓(xùn)練,并預(yù)測網(wǎng)絡(luò)時延,與LSTM神經(jīng)網(wǎng)絡(luò)預(yù)測結(jié)果進行對比。

3.2 LSTM的仿真

將300組時延數(shù)據(jù)劃分為2組,取250組時延數(shù)據(jù)作為LSTM的訓(xùn)練樣本,50組數(shù)據(jù)作為LSTM的測試樣本,應(yīng)不超過訓(xùn)練樣本數(shù)200,故先將LSTM神經(jīng)網(wǎng)絡(luò)隱含層節(jié)點數(shù)n設(shè)為100,初始學(xué)習(xí)率ε設(shè)置為0.005,迭代次數(shù)為500,同時設(shè)置了LearnRateDropPeriod為250,LearnRateDropFactor為0.5,令學(xué)習(xí)率在250次迭代時下降到初始學(xué)習(xí)率的1/2,從而加快LSTM神經(jīng)網(wǎng)絡(luò)擬合速度,這是處理神經(jīng)網(wǎng)絡(luò)數(shù)據(jù)訓(xùn)練時的常用手段。首先采用200組訓(xùn)練數(shù)據(jù)進行LSTM模型訓(xùn)練,在訓(xùn)練好的模型上迭代輸出后50步的網(wǎng)絡(luò)時延數(shù)值。LSTM對網(wǎng)絡(luò)時延的預(yù)測如圖5所示,LSTM訓(xùn)練迭代次數(shù)與誤差的關(guān)系如圖6所示。

3.3 WOA-LSTM的仿真

采用鯨魚算法優(yōu)化后的LSTM對網(wǎng)絡(luò)時延進行預(yù)測,采用MAPE作為鯨魚算法的損失函數(shù),采用WOA算法優(yōu)化LSTM的學(xué)習(xí)率和神經(jīng)網(wǎng)絡(luò)隱含層神經(jīng)元個數(shù),鯨魚算法初始化種群選為10,迭代次數(shù)為500次,初始化參數(shù)(n,ε)的取值范圍是[100,200]和[0.001,0.01]。采用200組數(shù)據(jù)進行WOA-LSTM訓(xùn)練,確定最優(yōu)的隱含層神經(jīng)元個數(shù)n和學(xué)習(xí)率ε,利用訓(xùn)練好的WOA-LSTM預(yù)測后50個時間步長的網(wǎng)絡(luò)時延數(shù)值。WOA-LSTM對時延的預(yù)測如圖7所示,WOA-LSTM訓(xùn)練迭代次數(shù)與誤差的關(guān)系如圖8所示。

3.4 BP神經(jīng)網(wǎng)絡(luò)的仿真

同樣選用前250組時延數(shù)據(jù)作為訓(xùn)練樣本,輸入層神經(jīng)元個數(shù)為1個,輸出層為1個,即用上一時間步的值預(yù)測下一時間步的數(shù)值,50組作為測試樣本,前向傳輸預(yù)測結(jié)果,后向反饋損失函數(shù)的誤差不斷調(diào)整權(quán)值和閾值,從而將網(wǎng)絡(luò)結(jié)構(gòu)穩(wěn)定完成預(yù)測。由經(jīng)驗公式η=m+n+l,m,n分別為輸入輸出層節(jié)點個數(shù),l?。?~9)之間隨機整數(shù),則BP神經(jīng)網(wǎng)絡(luò)隱含層神經(jīng)元個數(shù)取10。學(xué)習(xí)率與LSTM神經(jīng)網(wǎng)絡(luò)保持一致取0.005,同樣地,訓(xùn)練時最大迭代次數(shù)設(shè)置為500。BP神經(jīng)網(wǎng)絡(luò)對時延的預(yù)測如圖9所示,BP神經(jīng)網(wǎng)絡(luò)訓(xùn)練迭代次數(shù)與誤差的關(guān)系如圖10所示。

如表1所示,LSTM和WOA-LSTM預(yù)測數(shù)據(jù)的均方根誤差RMSE分別為2.529和2.152 9,說明WOA優(yōu)化的LSTM在一定程度上提高了網(wǎng)絡(luò)時延預(yù)測誤差。其次,從圖6和圖8可知,LSTM預(yù)測模型誤差在第450次迭代的時候才發(fā)生收斂,WOA-LSTM神經(jīng)網(wǎng)絡(luò)預(yù)測模型得到最優(yōu)隱含層神經(jīng)元個數(shù)128和最優(yōu)學(xué)習(xí)率0.003 3,此時神經(jīng)網(wǎng)絡(luò)經(jīng)訓(xùn)練之后,在300次的時候已經(jīng)開始慢慢收斂,在400次附近迭代的時候,預(yù)測誤差基本上無太大變化。如圖10所示,BP神經(jīng)網(wǎng)絡(luò)均方根誤差為10.200 3,LSTM神經(jīng)網(wǎng)絡(luò)作為時間相關(guān)性較強的神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu),相比BP神經(jīng)網(wǎng)絡(luò)在時延預(yù)測方面準確性更高,而經(jīng)過鯨魚優(yōu)化算法(WOA)優(yōu)化的LSTM神經(jīng)網(wǎng)絡(luò)預(yù)測結(jié)果的誤差得到了進一步減小,但如圖10所示,BP神經(jīng)網(wǎng)絡(luò)網(wǎng)絡(luò)結(jié)構(gòu)與運算方式更加簡單,誤差達到收斂時的迭代次數(shù)很少。

4 結(jié) 語

以對流層散射通信鏈路組成的窄帶通信網(wǎng)網(wǎng)絡(luò)時延數(shù)據(jù)為基礎(chǔ),采用WOA-LSTM算法對時延數(shù)據(jù)進行預(yù)測,更好地為通信網(wǎng)絡(luò)鏈路選擇及網(wǎng)絡(luò)協(xié)議切換提供數(shù)據(jù)支持。

1)LSTM神經(jīng)網(wǎng)絡(luò)在預(yù)測通信網(wǎng)時延數(shù)據(jù)這種時間序列的數(shù)據(jù)時,相較于BP神經(jīng)網(wǎng)絡(luò)更具優(yōu)勢。

2)利用WOA優(yōu)化參數(shù)后的LSTM神經(jīng)網(wǎng)絡(luò)相比于LSTM神經(jīng)網(wǎng)絡(luò)能夠更好地預(yù)測窄帶通信網(wǎng)的時延,預(yù)測精度提高了14.87%,誤差精度達到收斂時算法迭代次數(shù)更少,預(yù)測精度更高。

3)基于WOA-LSTM網(wǎng)絡(luò)時延預(yù)測算法預(yù)測精度相較于LSTM和BP神經(jīng)網(wǎng)絡(luò)算法更好,WOA-LSTM算法達到收斂時迭代次數(shù)相較于LSTM神經(jīng)網(wǎng)絡(luò)算法更少。

因此,本文提出的基于WOA-LSTM神經(jīng)網(wǎng)絡(luò)的網(wǎng)絡(luò)時延預(yù)測算法,具有較高的預(yù)測精度。但此預(yù)測算法的迭代速度有待進一步優(yōu)化,且只適用于本文所采集的數(shù)據(jù)類型,下一步將通過WOA-LSTM神經(jīng)網(wǎng)絡(luò)對其他通信手段組成的窄帶通信網(wǎng)網(wǎng)絡(luò)時延數(shù)據(jù)進行預(yù)測,從而探索此算法的適用范圍。

參考文獻/References:

[1] 徐旺,葛愿,王炎.基于ARIMA的NCS隨機時延預(yù)測[J].安徽工程大學(xué)學(xué)報,2016,31(4):72-76.

XU Wang,GE Yuan,WANG Yan.Predicting NCS stochastic delay based on ARIMA[J].Journal of Anhui Polytechnic University,2016,31(4):72-76.

[2] 汪知宇,張彤.基于改進LS-SVM算法的列車通信網(wǎng)絡(luò)時延預(yù)測方法[J].城市軌道交通研究,2021,24(1):101-106.

WANG Zhiyu,ZHANG Tong.Time delay prediction method for train communication network based on improved LS-SVM algorithm[J].Urban Mass Transit,2021,24(1):101-106.

[3] 時維國,雷何芬.基于PSO-BP神經(jīng)網(wǎng)絡(luò)的網(wǎng)絡(luò)時延預(yù)測算法[J].自動化與儀表,2020,35(7):1-5.

SHI Weiguo,LEI Hefen .Algorithm prediction of network delay using BP neural network based on particle swarm optimization[J].Automation & Instrumentation,2020,35(7):1-5.

[4] 法比奧,艾倫.神經(jīng)網(wǎng)絡(luò)算法與實現(xiàn)[M].北京:人民郵電出版社,2017.

[5] 張冬雯,趙琪,許云峰,等.基于長短期記憶神經(jīng)網(wǎng)絡(luò)模型的空氣質(zhì)量預(yù)測[J].河北科技大學(xué)學(xué)報,2020,41(1):66-75.

ZHANG Dongwen,ZHAO Qi,XU Yunfeng,et al.Air quality prediction based on neural network model of long short-term memory[J].Journal of Hebei University of Science and Technology,2020,41(1):66-75.

[6] 李勃.基于LSTM法的高速公路邊坡穩(wěn)定性研究[J].河北工業(yè)科技,2021,38(2):142-147.

LI Bo.Research on highway slope stability based on LSTM method[J].Hebei Journal of Industrial Science and Technology,2021,38(2):142-147.

[7] 張萍,肖為周,沈錚璽.基于長短期記憶網(wǎng)絡(luò)的軌道交通短期OD客流量預(yù)測[J].河北工業(yè)科技,2021,38(5):351-356.

ZHANG Ping,XIAO Weizhou,SHEN Zhengxi.Forecast of short-term origin-destination passenger flow of rail Transit based on long short-term memory network[J].Hebei Journal of Industrial Science and Technology,2021,38(5):351-356.

[8] SAID A B,ERRADI A,ALY H A,et al.Predicting COVID-19 cases using bidirectional LSTM on multivariate time series[J].Environmental Science and Pollution Research International,2021,28(40):56043-56052.

[9] GUO Aixia,BEHESHTI R,KHAN Y M,et al.Predicting cardiovascular health trajectories in time-series electronic health records with LSTM models[J].BMC Medical Informatics and Decision Making,2021,21(1):5.

[10] 張蕾,孫尚紅,王月.基于深度學(xué)習(xí)LSTM模型的匯率預(yù)測[J].統(tǒng)計與決策,2021,37(13):158-162.

ZHANG Lei,SUN Shanghong,WANG Yue.Exchange rate prediction based on deep learning LSTM model[J].Statistics and Decision,2021,37(13):158-162.

[11] 丁文絹.基于股票預(yù)測的ARIMA模型、LSTM模型比較[J].工業(yè)控制計算機,2021,34(7):109-112.

DING Wenjuan.Comparison of ARIMA model and LSTM model based on stock forecast[J].Industrial Control Computer,2021,34(7):109-112.

[12] 田聰.基于改進型EMD-LSTM的高頻金融時間序列預(yù)測[D].南昌:江西財經(jīng)大學(xué),2021.

TIAN Cong.High-Frequency Financial Time Series Prediction Based on Improved EMD-LSTM[D].Nanchang:Jiangxi University of Finance and Economics,2021.

[13] 鄭羅春.基于LSTM網(wǎng)絡(luò)模型的高速公路軟基長期沉降預(yù)測[J].湖南交通科技,2021,47(2):94-97.

ZHENG Luochun.Long-term settlement prediction of expressway soft foundation based on LSTM network model[J].Hunan Communication Science and Technology,2021,47(2):94-97.

[14] MIRJALILI S,LEWIS A.The whale optimization algorithm[J].Advances in Engineering Software,2016,95:51-67.

[15] 江旭東.基于EEMD-WOA-LSTM的電力負荷能耗預(yù)測系統(tǒng)的設(shè)計與實現(xiàn)[D].天津:天津理工大學(xué),2021.

JIANG Xudong.Design and Implementation of Power Load Energy Consumption Forecasting System Based on EEMD-WOA-LSTM[D].Tianjin:Tianjin University of Technology,2021.

[16] 李卓漫,王海瑞.基于PSO優(yōu)化LSTM的滾動軸承剩余壽命預(yù)測[J].化工自動化及儀表,2021,48(4):353-357.

LI Zhuoman,WANG Hairui .Predicting the remaining service Life of rolling bearings based on PSO-LSTM network model[J].Control and Instruments in Chemical Industry,2021,48(4):353-357.

宿松县| 东阿县| 鲁甸县| 札达县| 恩平市| 四平市| 焉耆| 安岳县| 资阳市| 江西省| 南溪县| 慈溪市| 蓬溪县| 长治市| 瑞安市| 聊城市| 濮阳县| 汕头市| 黎川县| 泰和县| 虹口区| 九江市| 伊春市| 张家川| 寿阳县| 庄浪县| 井研县| 东丰县| 聂拉木县| 濮阳市| 旬阳县| 长兴县| 晋城| 扬州市| 丰城市| 贺州市| 镇坪县| 久治县| 吉木乃县| 尚义县| 中宁县|