孫曉婷+劉年?yáng)|+杜坤+周明+任剛紅
摘 要:城市供水量是非線(xiàn)性、非平穩(wěn)時(shí)間序列,組合預(yù)測(cè)模型能獲得更高精度預(yù)測(cè)結(jié)果。通過(guò)深入分析混沌局域法與神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)模型特點(diǎn),提出了一種新的組合預(yù)測(cè)模型。首先,應(yīng)用混沌局域法對(duì)城市日供水量進(jìn)行初預(yù)測(cè),然后,應(yīng)用神經(jīng)網(wǎng)絡(luò)對(duì)預(yù)測(cè)結(jié)果進(jìn)行修正。由于所提出的組合模型利用了混沌局域法及神經(jīng)網(wǎng)絡(luò)進(jìn)行優(yōu)勢(shì)互補(bǔ),能同時(shí)提高預(yù)測(cè)精度與計(jì)算效率。為驗(yàn)證所提出組合預(yù)測(cè)模型的可行性,采用某市7 a實(shí)測(cè)供水量數(shù)據(jù),對(duì)混沌局域法、BPNN、RBF及GRNN神經(jīng)網(wǎng)絡(luò)4種單一預(yù)測(cè)模型及相應(yīng)的3種組合模型預(yù)測(cè)精度進(jìn)行定量分析,結(jié)果表明,組合預(yù)測(cè)模型精度都高于對(duì)應(yīng)單一預(yù)測(cè)模型,混沌局域法與GRNN神經(jīng)網(wǎng)絡(luò)組合模型預(yù)測(cè)精度最高,且運(yùn)算時(shí)間遠(yuǎn)低于單一神經(jīng)網(wǎng)絡(luò)模型運(yùn)算時(shí)間。
關(guān)鍵詞:混沌局域法;神經(jīng)網(wǎng)絡(luò);組合模型;日供水量預(yù)測(cè)
中圖分類(lèi)號(hào):TP183
文獻(xiàn)標(biāo)志碼:A 文章編號(hào):1674-4764(2017)05-0135-05
Abstract:Urban water supply is a nonlinear and non-stationary time series, and the combination forecasting model can get more accurate results. Through in-depth analysis of chaotic local-region method and neural network prediction model, this paper puts forward a new combination forecasting model, which uses chaotic local-region method to make a preliminary forecast for urban daily water supply, and then the prediction result is updated by neural network. The proposed combined model makes use of complementary advantages of the chaotic local-region method and the neural network, improving synchronously the accuracy and computational efficiency of the prediction results. To verify the proposed model, the prediction accuracy of the four single prediction models of Chaotic local-region method,BPNN, RBF and GRNN neural network and three corresponding combined models are analyzed quantitatively using seven years water supply data. The results show that combination forecasting model is of higher accuracy than single prediction model, and chaotic local-region method plus GRNN neural network combination model has highest accuracy with much lower computation time than single neural network predication model.
Keywords:chaotic local-region method; neural network; combination model; daily water supply forecast
城市供水量預(yù)測(cè)能輔助供水調(diào)度,提高水廠(chǎng)管理水平與生產(chǎn)效率,一直是學(xué)者們關(guān)注的重點(diǎn)課題[1-2]。供水量預(yù)測(cè)模型可分為傳統(tǒng)預(yù)測(cè)模型和基于新技術(shù)預(yù)測(cè)模型[3],傳統(tǒng)模型需對(duì)數(shù)據(jù)序列性質(zhì)進(jìn)行假設(shè),例如,平穩(wěn)性假設(shè)或周期性假設(shè),若假設(shè)不合理,得出的預(yù)測(cè)模型則會(huì)嚴(yán)重失真;基于新技術(shù)的預(yù)測(cè)模型通過(guò)非線(xiàn)性、自適應(yīng)學(xué)習(xí)方法構(gòu)建模型,能克服傳統(tǒng)預(yù)測(cè)模型缺點(diǎn)。如Tiwari等[4]提出了一種基于小波技術(shù)的神經(jīng)網(wǎng)絡(luò)供水量短期預(yù)測(cè)模型,結(jié)果表明,其預(yù)測(cè)精度比傳統(tǒng)ARIMA、ARIMAX和WNN方法高。Bai等[5]分析了供水量序列的混沌特性,利用自適應(yīng)混沌粒子群優(yōu)化RVM模型參數(shù),提出一種多尺度的RVM供水量預(yù)測(cè)組合模型。陳敏等[6]根據(jù)混沌理論計(jì)算重構(gòu)相空間嵌入維數(shù),用于確定BP神經(jīng)網(wǎng)絡(luò)隱藏層節(jié)點(diǎn)個(gè)數(shù),提高了預(yù)測(cè)精度。
從信息利用角度來(lái)看,單一預(yù)測(cè)模型只能利用數(shù)據(jù)部分有效信息,僅能從一個(gè)側(cè)面刻畫(huà)數(shù)據(jù)序列規(guī)律,具有一定局限性;組合預(yù)測(cè)模型通過(guò)優(yōu)勢(shì)互補(bǔ),能更大程度挖掘數(shù)據(jù)信息,可望獲得更高精度預(yù)測(cè)結(jié)果。神經(jīng)網(wǎng)絡(luò)與混沌理論模型作為目前最廣泛使用的兩種新技術(shù)預(yù)測(cè)模型,都展現(xiàn)出較高的預(yù)測(cè)精度,但針對(duì)二者的組合預(yù)測(cè)模型研究較少。通過(guò)深入分析,筆者發(fā)現(xiàn),基于混沌理論的預(yù)測(cè)模型[7],例如:全域法、局域法等,其特點(diǎn)是能在海量數(shù)據(jù)樣本中迅速挖掘時(shí)間序列總體趨勢(shì),但對(duì)局部細(xì)節(jié)預(yù)測(cè)能力較差。相較而言,神經(jīng)網(wǎng)絡(luò)模型[8-9]非線(xiàn)性擬合能力更強(qiáng),能更準(zhǔn)確追蹤局部細(xì)節(jié),但當(dāng)樣本數(shù)據(jù)量較大時(shí),神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)模型存在訓(xùn)練時(shí)間長(zhǎng)、收斂慢且預(yù)測(cè)結(jié)果不確定的缺點(diǎn)。鑒于此,提出了一種新的組合預(yù)測(cè)方法,首先,利用混沌預(yù)測(cè)模型對(duì)數(shù)據(jù)序列總體趨勢(shì)進(jìn)行初預(yù)測(cè),然后,應(yīng)用小樣本訓(xùn)練神經(jīng)網(wǎng)絡(luò)對(duì)預(yù)測(cè)結(jié)果進(jìn)行局部修正,實(shí)現(xiàn)二者的優(yōu)勢(shì)互補(bǔ),進(jìn)而能同時(shí)提高預(yù)測(cè)結(jié)果精度與計(jì)算效率。考慮混沌局域模型能較好處理實(shí)際中存在噪聲的預(yù)測(cè)問(wèn)題,采用加權(quán)一階局域法進(jìn)行初預(yù)測(cè),重點(diǎn)研究了其與不同類(lèi)型神經(jīng)網(wǎng)絡(luò)的組合預(yù)測(cè)效果,并采用某市的實(shí)測(cè)供水量數(shù)據(jù)對(duì)單一模型及組合模型的預(yù)測(cè)精度、運(yùn)算時(shí)間進(jìn)行定量分析。endprint
1 基于混沌理論的加權(quán)一階局域預(yù)測(cè)法
2 神經(jīng)網(wǎng)絡(luò)及其與加權(quán)一階局域法的
組合預(yù)測(cè)模型
人工神經(jīng)網(wǎng)絡(luò)是一種能對(duì)信息進(jìn)行分布式并行處理的數(shù)學(xué)模型,其最大特點(diǎn)是具有自學(xué)習(xí)和自適應(yīng)能力[13]。根據(jù)已有文獻(xiàn),BP、RBF及GRNN神經(jīng)網(wǎng)絡(luò)發(fā)展相對(duì)成熟,被廣泛用于解決各類(lèi)預(yù)測(cè)問(wèn)題。BP神經(jīng)網(wǎng)絡(luò)[14-15],即反向傳播神經(jīng)網(wǎng)絡(luò),采用輸出與輸入之差逆向反饋校正的方法使實(shí)際輸出不斷逼近期望輸出。它也是是目前理論最為完備的神經(jīng)網(wǎng)絡(luò),但它要事先確定網(wǎng)絡(luò)的結(jié)構(gòu),參數(shù)調(diào)整復(fù)雜,人為干擾因素較強(qiáng),訓(xùn)練好的網(wǎng)絡(luò),在給它新的時(shí)序時(shí),就需要重新訓(xùn)練。RBF網(wǎng)絡(luò)[16],即徑向基神經(jīng)網(wǎng)絡(luò),利用用RBF作為隱單元的“基”構(gòu)成隱藏空間,將低維的模型輸入數(shù)據(jù)變換到高維空間內(nèi),使得在低維空間的線(xiàn)性不可分問(wèn)題在高維空間線(xiàn)性可分。廣義回歸神經(jīng)網(wǎng)絡(luò)[17-18](GRNN)也是一種徑向基神經(jīng)網(wǎng)絡(luò),僅需要調(diào)節(jié)一個(gè)參數(shù),因此,人為干擾因素較小。在小數(shù)據(jù)量的情況下,預(yù)測(cè)效果也很好且運(yùn)算時(shí)間更短。
3 案例分析
選取某市水廠(chǎng)2000年1月至2006年12月的日供水量驗(yàn)證預(yù)測(cè)模型精度。為消除年供水量時(shí)間序列的季節(jié)性和趨勢(shì)性、減少噪聲影響[19],選取2000~2006年1月的217個(gè)日供水量作為時(shí)間序列樣本,其中210個(gè)日供水量數(shù)據(jù)作為參考樣本,7個(gè)日供水量數(shù)據(jù)作為驗(yàn)證樣本,對(duì)于其它月份的預(yù)測(cè)可按此方法進(jìn)行處理。采用互信息法計(jì)算得τ=7,根據(jù)文獻(xiàn)[20]推薦的多嵌入維法計(jì)算得m=10,再依據(jù)BIC信息準(zhǔn)則選取鄰近相點(diǎn)K=7。首先采用加權(quán)一階局域預(yù)測(cè)法、BP、RBF及GRNN神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)模型對(duì)日供水量進(jìn)行單獨(dú)預(yù)測(cè),然而將3種神經(jīng)網(wǎng)絡(luò)分別與加權(quán)一階局域法進(jìn)行組合預(yù)測(cè),總體預(yù)測(cè)趨勢(shì)見(jiàn)圖2,局部細(xì)節(jié)見(jiàn)圖3。
如圖2、圖3所示,7種方法都能較好預(yù)測(cè)日供水量總體趨勢(shì),而對(duì)局部細(xì)節(jié)預(yù)測(cè)有一定差異。對(duì)預(yù)測(cè)結(jié)果按精度排序,將平均絕對(duì)誤差(MAPE)最小的置于首位,如表1所示。混沌局域法與GRNN組合模型預(yù)測(cè)精度最高,單一的RBF神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)精度最低。此外,組合預(yù)測(cè)模型的精度都高于相應(yīng)的單一模型,由于采用小樣本訓(xùn)練神經(jīng)網(wǎng)絡(luò),其運(yùn)算時(shí)間僅為3 s,達(dá)到了同時(shí)提高預(yù)測(cè)結(jié)果精度及計(jì)算效率的目的。
4 結(jié)論
供水量預(yù)測(cè)對(duì)提高水廠(chǎng)管理水平具有重要意義,本文提出了一種新的組合預(yù)測(cè)方法以提高供水量預(yù)測(cè)精度,其利用混沌加權(quán)一階局域法對(duì)供水量進(jìn)行初預(yù)測(cè),然后利用神經(jīng)網(wǎng)絡(luò)對(duì)預(yù)測(cè)結(jié)果進(jìn)行修正。采用某市7年實(shí)測(cè)的供水量數(shù)據(jù)對(duì)不同模型預(yù)測(cè)結(jié)果進(jìn)行評(píng)判,得到了如下結(jié)論:
1)對(duì)單一模型,神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)精度普遍高于混沌加權(quán)一階局域法,其中GRNN神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)精度最高,BP神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)精度次之,但神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)模型存在訓(xùn)練時(shí)間長(zhǎng)、預(yù)測(cè)結(jié)果不確定的缺點(diǎn)。
2)組合預(yù)測(cè)模型精度都高于對(duì)應(yīng)單一預(yù)測(cè)模型,其中加權(quán)一階局域法與GRNN神經(jīng)網(wǎng)絡(luò)的組合模型預(yù)測(cè)精度最高,加權(quán)一階局域法與RBF神經(jīng)網(wǎng)絡(luò)的組合模型預(yù)測(cè)精度最低,但仍高于所有單一預(yù)測(cè)模型。由于采用了小樣本訓(xùn)練神經(jīng)網(wǎng)絡(luò)模型,其運(yùn)算時(shí)間短且精度更高。
參考文獻(xiàn):
[1] YASAR A, BILGILI M, SIMSEK E. Water demand forecasting based on stepwise multiple nonlinear regression analysis [J]. Arabian Journal for Science and Engineering, 2012, 37(8): 2333-2341.
[2] 孫月峰,閆雅飛,張表志.基于T-S模型的模糊神經(jīng)網(wǎng)絡(luò)城市需水量預(yù)測(cè)方法研究[J].安全與環(huán)境學(xué)報(bào), 2013, 13(2): 136-139.
SUN Y F, YAN Y F, ZHANG B Z. Fuzzy neural network method based on the T-S model for the urban water demand forecasting and its application [J]. Journal of Safety and Environment, 2013,13(2): 136-139.(in Chinese)
[3] PAI Y, PU W, CHUAN L, et al. Dynamic forecast of daily urban water consumption using variable-structure support vector regression model[J]. Journal of Water Resource Planning and Management, 2014, 140: 1943-5452.
[4] TIWARI M K, ADAMOWSKI J. Urban water demand forecasting and uncertainty assessment using ensemble wavelet-bootstrap-neural network models [J]. Water Resources Research, 2013, 49(10): 6486-6507.
[5] BAI Y, WANG P, LI C,et al. A multi-scale relevance vector regression approach for daily urban water demand forecasting [J]. Journal of Hydrology, 2014, 517: 236-245.
[6] 陳敏,鄧英偉.基于混沌理論的城市用水量預(yù)測(cè)研究[J].計(jì)算技術(shù)與自動(dòng)化,2009,28(1):88-90.endprint
CHEN M, DEN Y W. Forecast research of urban water consumption based on chaos theory [J].Computing Technology and Automation, 2009,28(1):88-90.(in Chinese)
[7] 韓敏.混沌時(shí)間序列預(yù)測(cè)理論與方法[M].北京:中國(guó)水利水電出版社,2007:28-39.
HAN M. Prediction theory and method of chaotic time series [M].Beijing: China Water & Power Press, 2007:28-39.(in Chinese)
[8] 史峰,王小川,郁磊.MATLAB神經(jīng)網(wǎng)絡(luò)30個(gè)案例分析[M].北京:北京航空航天大學(xué)出版社,2011:73-80.
SHI F, WANG X C, YU L. 30 case analysis of MATLAB neural network [M].Beijing: Beihang University Press,2011:73-80.(in Chinese)
[9] 林海軍,滕召勝,遲海,等. 基于多RBF神經(jīng)網(wǎng)絡(luò)的汽車(chē)衡誤差補(bǔ)償[J].湖南大學(xué)學(xué)報(bào)(自然科學(xué)版),2010,37(5):40-44.
LIN H J, TENG Z S, CHI H, et al. Error compensation for truck scale based on multiple RBF neural networks[J].Journal of Hunan University (Natural Sciences), 2010,37(5):40-44.(in Chinese)
[10] TAKENS F. Determining strange attractors in turbulence [J]. Lecture Notes in Mathematics, 1981, 898: 361-381.
[11] 劉年?yáng)|,杜坤,周明,等.局域法鄰近點(diǎn)選取對(duì)降雨量預(yù)測(cè)精度影響研究[J].給水排水, 2016(Sup1):285-288.
LIU N D, DU K, ZHOU M, et al. Study on the influence of select the local-region method nearest neighbours on rainfall forecasting accuracy[J]. Water & Wastewater and Drainage, 2016(Sup1):285-288.(in Chinese)
[12] CARLOS C, MUOZ-COLLAZOS H, URBANO-MUOZ R. Clustering of web search results based on the cuckoo search algorithm and Balanced Bayesian Information Criterion [J]. Information Sciences, 2014, 281:248-264.
[13] TAORMINA R, CHAU K, SETHI R. Artificial neural network simulation of hourly 4 groundwater levels in a coastal aquifer system of the 5 Venice lagoon [J]. Engineering Applications of Artificial Intelligence, 2012, 25 (8): 1670-1676.
[14] KARSOLIYA S. Approximating number of hidden layer neurons in multiple hidden layer BPNN architecture [J]. International Journal of Engineering Trends and Technology, 2012, 3(6): 714-717.
[15] 戴張俊,余飛,陳善雄,等.淮北平原地基分層與承載力的BP網(wǎng)絡(luò)預(yù)測(cè)方法[J].土木建筑與環(huán)境工程, 2013,35(3) :18-24.
DAI Z J, YU F, CHEN S X, et al. Stratification and bearing capacity prediction method based on BP neural network for foundation in Huaibei Plain [J]. Journal of Civil, Architectural & Environmental Engineering, 2013,35(3) :18-24. (in Chinese)
[16] TSEKOURASA G E, TSIMIKASB J.On training RBF neural networks using input-output fuzzy clustering and particle swarm optimization [J]. Science Direct Fuzzy Sets and Systems,2013, 221:65-89.
[17] 戴虹,錢(qián)晉武,張震,等. GRNN在肌電預(yù)測(cè)踝關(guān)節(jié)運(yùn)動(dòng)中的應(yīng)用[J].儀器儀表學(xué)報(bào), 2013,34(4):846-851.
DAI H, QIAN J W, ZHANG Z, et al. Application of GRNN in ankle movement prediction based on surface electromyography [J]. Chinese Journal of Scientific Instrument, 2013,34 (4):846-851.(in Chinese)
[18] FIRATA M,TURANB M E, YURDUSEV M A. Comparative analysis of neural network techniques for predicting water consumption time series [J]. Journal of Hydrology, 2010, 384 : 46-51.
[19] 趙瑩,盧文喜,羅建男,等.基于改進(jìn)時(shí)間序列分析法的鎮(zhèn)賚地區(qū)地下水位動(dòng)態(tài)分析[J].水利學(xué)報(bào),2013,44(11):1372-1379.
ZHAO Y,LU W X,LUO J N, et al. Dynamic analysis of groundwater table in Zhenlai county of Jilin province based on the improved time series analysis method [J]. Shuili Xuebao, 2013,44(11):1372-1379.(in Chinese)
[20] 劉道文,忽海娜.基于多嵌入維加權(quán)一階局域法煤礦災(zāi)害混沌時(shí)間序列預(yù)測(cè)[J].煤炭技術(shù),2012,31(8):89-91.
LIU D W, HU H N. Chaotic time series prediction of coal mine disasters based on multi-dimensional weighted first-order embedded local-region method [J]. Coal Technology, 2012,31(8):89-91. (in Chinese)
(編輯 胡玲)endprint