顧兆軍 李冰 劉濤
關(guān)鍵詞: 相空間重構(gòu); 粒子群算法; Elman神經(jīng)網(wǎng)絡(luò); 混沌時(shí)間序列; 網(wǎng)絡(luò)流量預(yù)測(cè); 參數(shù)優(yōu)化
中圖分類號(hào): TN98?34; TP393 ? ? ? ? ? ? ? ? ? ? 文獻(xiàn)標(biāo)識(shí)碼: A ? ? ? ? ? ? ? ? ? ? 文章編號(hào): 1004?373X(2019)01?0082?05
Abstract: The traditional neural network used for network traffic prediction is easy to fall into the local minimization, which may lead to the low prediction accuracy. Therefore, a network traffic prediction model is proposed on the basis of phase?space reconstruction, in which the particle swarm optimization (PSO) algorithm is used to optimize the initial parameter of Elman neural network. The phase?space reconstruction is carried out for the time sequence of network traffic, and then the reconstructed traffic sequence is taken as the input of the model. The global searching ability of PSO algorithm is utilized to optimize the initial parameter of Elman neural network. The trained Elman neural network is used to forecast the network traffic. The simulation results show that, in comparison with other traffic prediction models, the network prediction based on PSO?Elman model has higher prediction accuracy.
Keywords: phase?space reconstruction; PSO algorithm; Elman neural network; chaotic time series; network traffic prediction; parameter optimization
隨著互聯(lián)網(wǎng)技術(shù)的發(fā)展,網(wǎng)絡(luò)流量劇增,網(wǎng)絡(luò)阻塞問(wèn)題凸顯。建立準(zhǔn)確高效的網(wǎng)絡(luò)流量預(yù)測(cè)模型對(duì)異常流量的檢測(cè)、網(wǎng)絡(luò)規(guī)劃設(shè)計(jì)等都具有重要的意義。
網(wǎng)絡(luò)流量預(yù)測(cè)模型可分為線性和非線性兩種?;诙滔嚓P(guān)特性的自回歸模型[1]、自回歸滑動(dòng)平均模型[2]、自回歸綜合滑動(dòng)平均模型[3]以及基于長(zhǎng)相關(guān)特性的差分自回歸滑動(dòng)平均模型[4]均屬于線性預(yù)測(cè)模型。線性預(yù)測(cè)模型算法簡(jiǎn)單,易于實(shí)現(xiàn),但面對(duì)日益復(fù)雜的網(wǎng)絡(luò)流量,難以保證預(yù)測(cè)精度。針對(duì)網(wǎng)絡(luò)流量的突變性,非線性預(yù)測(cè)模型在流量預(yù)測(cè)中得到了較好的應(yīng)用。非線性預(yù)測(cè)模型主要包括灰色模型[5]、支持向量機(jī)[6]、小波預(yù)測(cè)模型[7]、神經(jīng)網(wǎng)絡(luò)[8?9]等。BP神經(jīng)網(wǎng)絡(luò)作為傳統(tǒng)神經(jīng)網(wǎng)絡(luò)的一種,憑借較強(qiáng)的非線性擬合能力、可映射任意復(fù)雜的非線性關(guān)系等特性,在流量預(yù)測(cè)領(lǐng)域得到了很好的應(yīng)用,但BP神經(jīng)網(wǎng)絡(luò)存在訓(xùn)練時(shí)間長(zhǎng)、易陷入局部極小等問(wèn)題[10]。Elman神經(jīng)網(wǎng)絡(luò)在傳統(tǒng)前饋神經(jīng)網(wǎng)絡(luò)的基礎(chǔ)上增加了有記憶功能的關(guān)聯(lián)層,使系統(tǒng)具有適應(yīng)時(shí)變特性的能力。但Elman算法的連接權(quán)值和隱含層參數(shù)隨機(jī)生成,在梯度下降算法調(diào)整時(shí)易陷入局部極小。PSO算法具有快速全局尋優(yōu)能力,利用PSO算法優(yōu)化Elman神經(jīng)網(wǎng)絡(luò)權(quán)值和閾值使其避免陷入局部極小,提高了網(wǎng)絡(luò)流量預(yù)測(cè)精度。同時(shí),網(wǎng)絡(luò)流量隨時(shí)間變化而變化,是典型的時(shí)間序列。研究表明,網(wǎng)絡(luò)流量具有混沌性和時(shí)變性[11],用混沌動(dòng)力學(xué)處理時(shí)間序列是一個(gè)較好的處理方式,提高了理論分析問(wèn)題的能力。本文提出在相空間重構(gòu)基礎(chǔ)上采用PSO算法優(yōu)化Elman神經(jīng)網(wǎng)絡(luò)模型的網(wǎng)絡(luò)流量預(yù)測(cè)模型,通過(guò)仿真實(shí)驗(yàn)證明了該模型的有效性。
通過(guò)對(duì)圖4,圖5及表1進(jìn)行分析,可得到如下結(jié)論:
1) 利用PSO算法對(duì)于Elman的初始參數(shù)進(jìn)行尋優(yōu),只需要迭代20次,結(jié)果就會(huì)收斂到一個(gè)穩(wěn)定的狀態(tài),預(yù)測(cè)精度即適應(yīng)度較迭代之前可以有了明顯的提高。
2) PSO?Elman算法預(yù)測(cè)序列走勢(shì)最為接近真實(shí)值,其他算法的擬合效果均比PSO?Elman差。
3) 通過(guò)表1可以發(fā)現(xiàn),對(duì)于RMSE,MAE,MRE三個(gè)指標(biāo),基于PSO?Elman神經(jīng)網(wǎng)絡(luò)的計(jì)算結(jié)果均遠(yuǎn)遠(yuǎn)小于其他兩種算法,證明本文算法的預(yù)測(cè)精度優(yōu)于其他算法。
本文提出一種網(wǎng)絡(luò)流量預(yù)測(cè)新方法。根據(jù)網(wǎng)絡(luò)流量混沌性和時(shí)變性特點(diǎn),用混沌時(shí)間序列處理網(wǎng)絡(luò)流量,提高了理論分析問(wèn)題的能力。同時(shí),本文采用的PSO?Elman算法不僅克服了傳統(tǒng)BP神經(jīng)網(wǎng)絡(luò)缺少關(guān)聯(lián)層導(dǎo)致系統(tǒng)沒(méi)有動(dòng)態(tài)記憶功能的缺點(diǎn),而且也克服了BP神經(jīng)網(wǎng)絡(luò)以及Elman神經(jīng)網(wǎng)絡(luò)易陷入局部極小、收斂速度慢的問(wèn)題,極大地提高了預(yù)測(cè)網(wǎng)絡(luò)流量的精度,有著良好的實(shí)踐意義。仿真結(jié)果表明,相比于BP神經(jīng)網(wǎng)絡(luò)、Elman算法,在相空間重構(gòu)的基礎(chǔ)上,采用PSO?Elman模型的網(wǎng)絡(luò)流量預(yù)測(cè)具有更高的準(zhǔn)確率。
參考文獻(xiàn)
[1] 金瑤,蔡之華.基于AR模型的Kalman濾波在股票價(jià)格預(yù)測(cè)中的應(yīng)用[J].統(tǒng)計(jì)與決策,2013,6(3):80?82.
JIN Yao, CAI Zhihua. Application of Kalman filter based on AR model in stock price forecasting [J]. Statistics & decision, 2013, 6(3): 80?82.
[2] 劉明風(fēng),修春波.基于ARMA與神經(jīng)網(wǎng)絡(luò)的風(fēng)速序列混合預(yù)測(cè)方法[J].中南大學(xué)學(xué)報(bào)(自然科學(xué)版),2013,44(1):16?20.
LIU Mingfeng, XIU Chunbo. Hybrid prediction method for wind speed series based on ARMA and neural network [J]. Journal of Central South University (natural science), 2013, 44(1): 16?20.
[3] 孫軼軒,邵春福,計(jì)尋,等.基于ARIMA與信息?;疭VR組合模型的交通事故時(shí)序預(yù)測(cè)[J].清華大學(xué)學(xué)報(bào)(自然科學(xué)版),2014,54(3):348?353.
SUN Yixuan, SHAO Chunfu, JI Xun, et al. Urban traffic accident time series prediction model based on combination of ARIMA and information granulation SVR [J]. Journal of Tsinghua University (natural science), 2014, 54(3): 348?353.
[4] 陳珂,彭志平,柯文德.基于人工免疫和FARIMA模型的流量預(yù)測(cè)方法研究[J].四川大學(xué)學(xué)報(bào)(自然科學(xué)版),2013,50(2):267?272.
CHEN Ke, PENG Zhiping, KE Wende. The study of traffic prediction method based on artificial immune and FARIMA model [J]. Journal of Sichuan University (natural science edition), 2013, 50(2): 267?272.
[5] 王相守,張娜.基于灰色神經(jīng)網(wǎng)絡(luò)組合模型的光伏短期出力預(yù)測(cè)[J].電力系統(tǒng)自動(dòng)化,2012,36(19):37?41.
WANG Xiangshou, ZHANG Na. Short?term output power forecast of photovoltaic based on a grey and neural network hybrid model [J]. Automation of electric power systems, 2012, 36(19): 37?41.
[6] 唐舟進(jìn),彭濤,王文博.一種基于相關(guān)分析的局域最小二乘支持向量機(jī)小尺度網(wǎng)絡(luò)流量預(yù)測(cè)算法[J].物理學(xué)報(bào),2014,63(13):1?10.
TANG Zhoujin, PENG Tao, WANG Wenbo. A local least square support vector machine prediction algorithm of small scale network traffic based on correlation analysis [J]. Acta physica Sinica, 2014, 63(13): 1?10.
[7] 李志農(nóng),朱明,褚福磊,等.基于經(jīng)驗(yàn)小波變換的機(jī)械故障診斷方法研究[J].儀器儀表學(xué)報(bào),2014,35(11):2423?2432.
LI Zhinong, ZHU Ming, CHU Fulei, et al. Mechanical fault diagnosis method based on empirical wavelet transform [J]. Chinese journal of scientific instrument, 2014, 35(11): 2423?2432.
[8] 郭通,蘭巨龍,李玉峰,等.基于量子自適應(yīng)粒子群優(yōu)化徑向基函數(shù)神經(jīng)網(wǎng)絡(luò)的網(wǎng)絡(luò)流量預(yù)測(cè)[J].電子與信息學(xué)報(bào),2013,35(9):2220?2226.
GUO Tong, LAN Julong, LI Yufeng, et al. Network traffic prediction with radial basis function neural network based on quantum adaptive particle swarm optimization [J]. Journal of electronics & information technology, 2013, 35(9): 2220?2226.
[9] 李翔,朱全銀.Adaboost算法改進(jìn)BP神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)研究神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)研究[J].計(jì)算機(jī)工程與科學(xué),2013,35(8):96?102.
LI Xiang, ZHU Quanyin. Prediction of improved BP neural network by Adaboost algorithm [J]. Computer engineering & science, 2013, 35(8): 96?102.
[10] 張立仿,張喜平.量子遺傳算法優(yōu)化BP神經(jīng)網(wǎng)絡(luò)的網(wǎng)絡(luò)流量預(yù)測(cè)[J].計(jì)算機(jī)工程與科學(xué),2016,38(1):114?119.
ZHANG Lifang, ZHANG Xiping. Network traffic prediction based on BP neural networks optimized by quantum genetic algorithm [J]. Computer engineering & science, 2016, 38(1): 114?119.
[11] LU Jinjun, WANG Zhiquan. Internet traffic data flow forecast by RFB neural network based on phase space reconstruction [J]. Transactions of Nanjing University of Aeronautics & Astronutics, 2006, 23(4): 52?63.
[12] Bellcore Lab. Traces available in the internet traffic archive [EB/OL]. [2007?03?24]. http://ita.ee.lbl.gov/html/traces.html.
[13] 姜明,吳春明,張旻,等.網(wǎng)絡(luò)流量預(yù)測(cè)中的時(shí)間序列模型比較研究[J].電子學(xué)報(bào),2009,37(11):2353?2358.
JIANG Ming, WU Chunming, ZHANG Min, et al. Research on the comparison of time series models for network traffic prediction [J]. Acta electronica Sinica, 2009, 37(11): 2353?2358.
[14] VAN DEN BERGH F. An analysis of particle swarm optimi?zers [D]. South Africa: University of Pretoria, 2001.