章 靜,叢騰龍,蘇光輝,秋穗正
(1.西安交通大學(xué)動(dòng)力工程多相流國(guó)家重點(diǎn)實(shí)驗(yàn)室,陜西西安 710049;2.西安交通大學(xué)核科學(xué)與技術(shù)學(xué)院,陜西西安 710049)
遺傳神經(jīng)網(wǎng)絡(luò)對(duì)水平通道流動(dòng)沸騰傳熱系數(shù)的預(yù)測(cè)
章 靜1,2,叢騰龍1,2,蘇光輝1,2,秋穗正1,2
(1.西安交通大學(xué)動(dòng)力工程多相流國(guó)家重點(diǎn)實(shí)驗(yàn)室,陜西西安 710049;2.西安交通大學(xué)核科學(xué)與技術(shù)學(xué)院,陜西西安 710049)
分別采用3層反向傳播神經(jīng)網(wǎng)絡(luò)(BPN)和遺傳神經(jīng)網(wǎng)絡(luò)(GNN)預(yù)測(cè)從常規(guī)通道到微通道尺度范圍內(nèi)的管內(nèi)流動(dòng)沸騰傳熱系數(shù),GNN的精度優(yōu)于BPN的精度(均方根誤差分別為17.16%和20.50%)。輸入?yún)?shù)為含氣率、質(zhì)量流密度、熱流密度、管徑和物性,輸出參數(shù)為傳熱系數(shù)。基于GNN預(yù)測(cè)結(jié)果,進(jìn)行了參數(shù)趨勢(shì)分析。對(duì)常規(guī)通道,傳熱系數(shù)隨壓力的增大而增大;對(duì)微通道,低壓時(shí)傳熱系數(shù)受壓力影響很小,高壓、低含氣率時(shí),傳熱系數(shù)隨壓力的增大而增大,高壓、高含氣率時(shí),傳熱系數(shù)隨壓力的增大而減小。傳熱系數(shù)隨質(zhì)量流密度、熱流密度的增大而增大。隨含氣率的增大,傳熱系數(shù)先增大后減?。晃⑼ǖ腊l(fā)生燒干時(shí)的含氣率較低。傳熱系數(shù)隨管徑的減小而增大;管徑越小,越易發(fā)生燒干。
BP神經(jīng)網(wǎng)絡(luò);遺傳神經(jīng)網(wǎng)絡(luò);流動(dòng)沸騰傳熱系數(shù)
管內(nèi)流動(dòng)沸騰傳熱系數(shù)一直是研究的熱點(diǎn)。核反應(yīng)堆中的傳熱為沸騰傳熱,能夠提高反應(yīng)堆系統(tǒng)的效率[1]。因此,準(zhǔn)確預(yù)測(cè)流動(dòng)沸騰傳熱系數(shù)、在安全的范圍內(nèi)盡可能的提高系統(tǒng)效率十分重要。近年來,人們對(duì)沸騰傳熱的研究從常規(guī)通道逐漸發(fā)展到了微通道。
對(duì)于常規(guī)通道,較常用的經(jīng)驗(yàn)關(guān)系式有Wattelet、Jung、Shah、Gungor-Winterton、Kandlikar公式[2-6]。由于常規(guī)通道和微通道流動(dòng)沸騰傳熱機(jī)理不同,以上公式不能很好地預(yù)測(cè)微通道的流動(dòng)沸騰傳熱系數(shù),且現(xiàn)有的微通道公式是在很窄的試驗(yàn)范圍內(nèi)擬合的,不能很好地預(yù)測(cè)較大工況范圍內(nèi)微通道的流動(dòng)沸騰傳熱系數(shù)。
人工神經(jīng)網(wǎng)絡(luò)有自適應(yīng)性、自組織性和很強(qiáng)的學(xué)習(xí)能力。遺傳神經(jīng)網(wǎng)絡(luò)(GNN)具有全局搜索的優(yōu)勢(shì)且收斂速度較快。Wang等[7]曾用神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)了水平常規(guī)通道傳熱系數(shù)。相比于神經(jīng)網(wǎng)絡(luò)方法,遺傳神經(jīng)網(wǎng)絡(luò)的預(yù)測(cè)結(jié)果更準(zhǔn)確。本文將運(yùn)用遺傳神經(jīng)網(wǎng)絡(luò)來預(yù)測(cè)常規(guī)通道到微通道的沸騰傳熱系數(shù)。
1.1 基本神經(jīng)網(wǎng)絡(luò)
人工神經(jīng)網(wǎng)絡(luò)是基于模擬生物神經(jīng)元的機(jī)理的計(jì)算結(jié)構(gòu),是生物神經(jīng)元的抽象和簡(jiǎn)化。神經(jīng)網(wǎng)絡(luò)的信息處理單元稱為神經(jīng)元,或稱為節(jié)點(diǎn)。
數(shù)學(xué)模型如下:其中:oj(t)為t時(shí)刻神經(jīng)元的信息輸入;τij為輸入輸出間的突觸延時(shí);Tj為神經(jīng)元j的閾值;wij為神經(jīng)元i到j(luò)的突觸連接系數(shù),即權(quán)重;f(·)為神經(jīng)元傳遞函數(shù)。
本文選擇反向傳播神經(jīng)網(wǎng)絡(luò)(BPN)來進(jìn)行分析。Balcilar等[8]認(rèn)為,BPN的誤差小、精度高,能很好地預(yù)測(cè)流動(dòng)沸騰傳熱系數(shù)。
1.2 遺傳神經(jīng)網(wǎng)絡(luò)
值得注意的是,BP神經(jīng)網(wǎng)絡(luò)有收斂速度過慢及易陷入局部極值點(diǎn)等缺點(diǎn),在實(shí)驗(yàn)數(shù)據(jù)有限的情況下這個(gè)缺點(diǎn)將尤其突出。遺傳神經(jīng)網(wǎng)絡(luò)利用遺傳算法優(yōu)化神經(jīng)網(wǎng)絡(luò)的閾值和權(quán)重,能減少實(shí)驗(yàn)的次數(shù)并提高神經(jīng)網(wǎng)絡(luò)的精度。其流程圖示于圖1,遺傳神經(jīng)網(wǎng)絡(luò)通過編碼、適應(yīng)度計(jì)算以及選擇交叉變異等遺傳操作,得到適合神經(jīng)網(wǎng)絡(luò)的權(quán)值和閾值,將所得的權(quán)值和閾值傳遞給神經(jīng)網(wǎng)絡(luò)。
圖1 遺傳神經(jīng)網(wǎng)絡(luò)流程圖Fig.1 Flow chart of GNN
2.1 訓(xùn)練神經(jīng)網(wǎng)絡(luò)
訓(xùn)練神經(jīng)網(wǎng)絡(luò)所使用的實(shí)驗(yàn)數(shù)據(jù)源于文獻(xiàn)[9-16],這些數(shù)據(jù)覆蓋了常規(guī)通道和微通道情況,實(shí)驗(yàn)參數(shù)范圍列于表1。
表1 訓(xùn)練神經(jīng)網(wǎng)絡(luò)所用參數(shù)范圍Table 1 Range of testing data for trained neural network
pf動(dòng)沸騰傳熱系數(shù)h。
神經(jīng)網(wǎng)絡(luò)的隱層神經(jīng)元數(shù)對(duì)網(wǎng)絡(luò)質(zhì)量的影響很大,目前尚無通用的隱層神經(jīng)元數(shù)預(yù)測(cè)方法[18],本文采用逐個(gè)嘗試的方法確定該值。文中分別將隱層神經(jīng)元數(shù)設(shè)置為5~30,為排除過擬合影響,對(duì)每種結(jié)構(gòu)進(jìn)行5次訓(xùn)練,去掉誤差最大、最小的結(jié)果,將剩余值求平均,得到該種結(jié)構(gòu)的誤差值,圖2為隱含層神經(jīng)元數(shù)對(duì)誤差的影響。由圖2可見,神經(jīng)網(wǎng)絡(luò)的誤差隨隱層神經(jīng)元的增加呈減小趨勢(shì),且減小趨勢(shì)逐漸減緩,在隱層神經(jīng)元數(shù)達(dá)22后,誤差減小幅度很小??紤]到過度增大隱層神經(jīng)元數(shù)會(huì)增加計(jì)算時(shí)間并導(dǎo)致過擬合,因此本文隱層神經(jīng)元數(shù)取為22。
圖2 隱含層神經(jīng)元數(shù)對(duì)神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)誤差的影響Fig.2 Influence of hidden layer neuron number on prediction errors of neural network
圖3為兩種神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)的傳熱系數(shù)與實(shí)驗(yàn)值的對(duì)比??煽闯?,BPN和GNN均能準(zhǔn)確預(yù)測(cè)傳熱系數(shù),且GNN的預(yù)測(cè)精度較BPN的高。兩種網(wǎng)絡(luò)的誤差及可信度分布列于表2和3。
圖3 BPN與GNN的計(jì)算結(jié)果與實(shí)驗(yàn)值的對(duì)比Fig.3 Comparison of predictions by BPN,GNN and experimental values
表2 不同模型預(yù)測(cè)誤差分析Table 2 Error analyses of BPN and GNN
表3 不同模型預(yù)測(cè)可信度分析Table 3 Confidences of BPN and GNN
2.2 各參數(shù)對(duì)傳熱系數(shù)的影響
本節(jié)基于GNN采用內(nèi)插預(yù)測(cè)結(jié)果分析各參數(shù)(p、x、G、q、d)對(duì)流動(dòng)沸騰傳熱系數(shù)的影響。
1)壓力的影響
壓力對(duì)流動(dòng)沸騰傳熱系數(shù)的影響如圖4所示。在常規(guī)通道中,傳熱系數(shù)隨壓力的增大而增大。這是因?yàn)閴毫υ龃笫贡砻鎻埩p小,從而使得受熱面能形成更多的汽化核心且能加快氣泡脫離壁面速率,由此促進(jìn)核態(tài)沸騰。壓力增大還能使氣液兩相的密度差減小,從而使液膜厚度增加。液膜厚度增加將導(dǎo)致壁面過熱度增加,進(jìn)而汽化核心增加,促進(jìn)核態(tài)沸騰;雖然質(zhì)量流密度一定時(shí),液膜厚度增加將會(huì)導(dǎo)致流體流速減小,從而輕微抑制對(duì)流沸騰,但對(duì)于常規(guī)通道,壓力對(duì)核態(tài)沸騰的促進(jìn)作用影響更大,因此傳熱系數(shù)隨壓力的增大,仍呈增大趨勢(shì)。
在微通道中,壓力較低(圖4c)時(shí)傳熱系數(shù)受壓力的影響較??;壓力較高(圖4d)時(shí),傳熱系數(shù)受壓力的影響大,這是因?yàn)閴毫?duì)核態(tài)沸騰起促進(jìn)作用。由圖4d可看出,當(dāng)含氣率較低時(shí),傳熱系數(shù)隨壓力的增大而增大。這是因?yàn)榈秃瑲饴蕰r(shí),核態(tài)沸騰起主導(dǎo)作用。當(dāng)含氣率較高時(shí),傳熱系數(shù)隨壓力的增大而減小,因?yàn)樵诟吆瑲饴氏?,?duì)流沸騰起主導(dǎo)作用,而壓力增大對(duì)對(duì)流沸騰起抑制作用。
圖4 壓力對(duì)傳熱系數(shù)的影響Fig.4 Influence of pressure on HTC
2)質(zhì)量流密度的影響
圖5為質(zhì)量流密度對(duì)流動(dòng)沸騰傳熱系數(shù)的影響。常規(guī)通道(圖5a、b)中,傳熱系數(shù)隨質(zhì)量流密度的增大而整體呈增大趨勢(shì),且含氣率越大,增速越大。這是因?yàn)樵龃筚|(zhì)量流密度對(duì)流沸騰有促進(jìn)作用,高含氣率情況下,對(duì)流作用更明顯。低質(zhì)量流密度(G<400kg/(m2·s))、低含氣率(x=0.2,0.3)時(shí)傳熱系數(shù)隨質(zhì)量流密度增大而減小(圖5b),這是因?yàn)樵黾淤|(zhì)量流量將導(dǎo)致界面剪切力增加,使得脫離壁面的氣泡直徑減小,從而輕微抑制核態(tài)沸騰[13]。低質(zhì)量流密度、低含氣率情況下,核態(tài)沸騰占主導(dǎo)地位(圖5b)。
微通道(圖5c)以核態(tài)沸騰為主,而傳熱系數(shù)受質(zhì)量流密度的影響很小。當(dāng)含氣率增大時(shí),對(duì)流沸騰影響加大,傳熱系數(shù)受質(zhì)量流密度的影響較大。
3)熱流密度的影響
熱流密度對(duì)流動(dòng)沸騰傳熱系數(shù)的影響如圖6所示。由圖6可見,無論是常規(guī)通道還是微通道,傳熱系數(shù)隨熱流密度的增加而增大。這是因?yàn)闊崃髅芏仍黾邮箽馀菝撾x壁面速度增加,并增加了汽化核心的數(shù)目,從而促進(jìn)了核態(tài)沸騰,熱流密度對(duì)對(duì)流沸騰則幾乎無影響[13]。
圖5 質(zhì)量流密度對(duì)傳熱系數(shù)的影響Fig.5 Influence of mass flux on HTC
圖6 熱流密度對(duì)傳熱系數(shù)的影響Fig.6 Influence of heat flux on HTC
常規(guī)通道(圖6a、b)中質(zhì)量流密度較低。由圖6a可看出,傳熱系數(shù)隨熱流密度的增大而增大,核態(tài)沸騰起作用;傳熱系數(shù)隨質(zhì)量流密度的增大而增大,這一現(xiàn)象由對(duì)流沸騰導(dǎo)致。該階段流動(dòng)沸騰傳熱中核態(tài)沸騰與對(duì)流沸騰共同作用。
微通道(圖6c、d)中,傳熱系數(shù)隨熱流密度的增大而大幅增大,說明處于核態(tài)沸騰主導(dǎo)階段。
4)含氣率的影響
圖7為含氣率對(duì)傳熱系數(shù)的影響。由圖7a可看出,常規(guī)通道中,傳熱系數(shù)隨含氣率的增大而增大。此時(shí),對(duì)流沸騰增大,核態(tài)沸騰被抑制。在圖中所示含氣率范圍內(nèi),未達(dá)到燒干,因此,并未出現(xiàn)傳熱系數(shù)陡降的情況。
在微通道中,低含氣率時(shí)傳熱系數(shù)受含氣率的影響較小,當(dāng)含氣率達(dá)一定值時(shí)出現(xiàn)突降。含氣率較低時(shí),傳熱系數(shù)卻同時(shí)受核態(tài)沸騰和對(duì)流沸騰兩種因素的影響,兩個(gè)作用相互抵消。因此,傳熱系數(shù)受含氣率的影響不大。微通道由于通道較小,燒干的情況早于常規(guī)通道出現(xiàn),此傳熱惡化導(dǎo)致傳熱系數(shù)的陡降。圖7b說明質(zhì)量流密度越大,傳熱系數(shù)降低發(fā)生得越早。
5)管徑的影響
圖8為管徑對(duì)傳熱系數(shù)的影響。常規(guī)通道中,傳熱系數(shù)隨管徑的增大而減小。管徑越小,單位體積加熱面越大。微通道中,管徑越小,發(fā)生燒干的情況越早越嚴(yán)重。
圖7 含氣率對(duì)傳熱系數(shù)的影響Fig.7 Influence of vapor quality on HTC
圖8 管徑對(duì)傳熱系數(shù)的影響Fig.8 Influence of diameter on HTC
本文分別訓(xùn)練了3層隱含層神經(jīng)元數(shù)為22的BPN和GNN,用于預(yù)測(cè)常規(guī)通道到微通道范圍內(nèi)的管內(nèi)對(duì)流沸騰傳熱系數(shù),并使用GNN進(jìn)行了參數(shù)影響分析,得到如下結(jié)論:
1)GNN和BPN預(yù)測(cè)的均方根誤差分別為17.16%和20.5%,兩者均能準(zhǔn)確預(yù)測(cè)傳熱系數(shù),且GNN預(yù)測(cè)的精度高于BPN。
2)微通道以核態(tài)沸騰為主,常規(guī)通道則以對(duì)流沸騰為主。
3)在常規(guī)通道中,傳熱系數(shù)隨壓力的增大而增大。對(duì)于微通道,當(dāng)含氣率較高時(shí),傳熱系數(shù)隨壓力的增大而減小。
4)傳熱系數(shù)隨質(zhì)量流密度和熱流密度的增大而增大,且含氣率越大,增速越大。
5)常規(guī)通道中,傳熱系數(shù)隨含氣率的增大而增大;在微通道中,傳熱系數(shù)隨含氣率的增加先增大后減小。
6)管徑越小,常規(guī)通道和微通道傳熱系數(shù)越高,微通道燒干惡化越嚴(yán)重。
[1] SU G H,F(xiàn)UKUD K,JIA D N,et al.Application of an artificial neural network in reactor thermohydraulic problem:Prediction of critical heat flux[J].Journal of Nuclear Science and Technology,2002,39(5):564-571.
[2] WATTELET J P,CHATO J C,SOUZA A L,et al.Evaporative characteristics of R-12,R-134a,and a mixture at low mass fluxes[J].ASHRAE Trans,1994,101:603-615.
[3] JUNG D,RADERMACHER R.Transport properties and surface tension of pure and mixed refrigerants[J].ASHRAE Trans,1991,97(1):90-99.
[4] SHAH M M.Chart correlation for saturated boiling heat transfer:Equations and further study[J].ASHRAE Trans,1982,88(1):185-196.
[5] GUNGOR K E,WINTERTON R H S.Simplified general correlation for saturated flow boiling and comparisons of correlations with data[J].Chemical Engineering Research &Design,1987,65(2):148-156.
[6] KANDLIKAR S G.A general correlation for saturated two-phase flow boiling heat transfer inside horizontal and vertical tubes[J].Journal of Heat Transfer,1990,112(1):219-228.
[7] WANG W J,ZHAO L X,ZHANG C L.Generalized neural network correlation for flow boiling heat transfer of R22and its alternative refrigerants inside horizontal smooth tubes[J].International Journal of Heat and Mass Transfer,2006,49(15):2 458-2 465.
[8] BALCILAR M,DALKILIC A S,WONGWISES S.Artificial neural network techniques for the determination of condensation heat transfer characteristics during downward annular flow of R134ainside a vertical smooth tube[J].Heat Mass Transfer,2011,38(1):75-84.
[9] GRECO A.Convective boiling of pure and mixed refrigerants:An experimental study of the major parameters affecting heat transfer[J].Heat Mass Transfer,2007,51(3-4):896-909.
[10]GRECO A,VANOLI G P.Evaporation of refrigerants in a smooth horizontal tube:Prediction of R22and R507heat transfer coefficients and pressure drop[J].Applied Thermal Engineering,2004,24(14-15):2 189-2 206.
[11]KUO C S.In-tube evaporation of HCFC-22in a 9.52mm micro-fin/smooth tube[J].Heat Mass Transfer,1996,39(12):2 559-2 569.
[12]WOJTAN L,URSENBACHER T,THOME J R.Investigation of flow boiling in horizontal tubes,PartⅠ:A new diabatic two-phase flow pattern map[J].Heat Mass Transfer,2005,48(14):2 955-2 969.
[13]WOJTAN L,URSENBACHER T,THOME J R.Investigation of flow boiling in horizontal tubes,PartⅡ:Development of a new heat transfer model for stratified-wavy,dryout and mist flow regimes[J].Heat Mass Transfer,2005,48(14):2 970-2 985.
[14]JUNG D,CHO Y,PARK K.Flow condensation heat transfer coefficients of R22,R134a,R407C,and R410Ainside plain and microfin tubes[J].International Journal of Refrigeration,2004,27(1):25-32.
[15]KIM Y,SEO K,CHUNG J T.Evaporation heat transfer characteristics of R-410Ain 7and 9.52 mm smooth/micro-fin tubes[J].International Journal of Refrigeration,2002,25(6):716-730.
[16]FILHOA E P B,JABARDOB J M S.Convective boiling performance of refrigerant R-134ain herringbone and microfin copper tubes[J].International Journal of Refrigeration,2006,29(1):81-91.
[17]張立明.人工神經(jīng)網(wǎng)絡(luò)的模型及其應(yīng)用[M].上海:復(fù)旦大學(xué)出版社,1993.
[18]SU G H,MORITA K,F(xiàn)UKUDA K,et al.Analysis of the critical heat flux in round vertical tubes under low pressure and flow oscillation conditions:Applications of artificial neural network[J].Nuclear Engineering and Design,2003,220(1):17-35.
Prediction of Flow Boiling Heat Transfer Coefficient in Horizontal Channel by Genetic Neural Network
ZHANG Jing1,2,CONG Teng-long1,2,SU Guang-hui1,2,QIU Sui-zheng1,2
(1.State Key Laboratory of Multiphase Flow in Power Engineering,Xi’an Jiaotong University,Xi’an710049,China;2.School of Nuclear Science and Technology,Xi’an Jiaotong University,Xi’an710049,China)
The three-layer back propagation network(BPN)and genetic neural network(GNN)were developed to predict the flow boiling heat transfer coefficient(HTC)in conventional and micro channels.The precision of GNN is higher than that of BPN(with root mean square errors of 17.16%and 20.50%,respectively).The inputs include vapor quality,mass flux,heat flux,diameter and physical properties and the output is HTC.Based on the trained GNN,the influences of input parameters on HTC were analyzed.HTC increases with pressure in conventional channels.The pressure has a negligible effect at low pressure region on HTC for micro channels.However,at high pressure region,HTC increases in low vapor quality region,while decreases in the highvapor quality region with the increase of pressure.HTC increases with the mass flux and heat flux,and HTC initially increases and then decreases as vapor quality increases.HTC increases inversely with the decrease of diameter.Dry-out arises at a lower quality in micro channels than that in conventional channels and more easily occurs in a smaller channel.
back propagation network;genetic neural network;flow boiling heat transfer coefficient
TL33
:A
:1000-6931(2015)01-0070-07
10.7538/yzk.2015.49.01.0070
2013-11-06;
2014-04-03
國(guó)家杰出青年科學(xué)基金資助項(xiàng)目(11125522)
章 靜(1989—),女,湖南衡陽人,博士研究生,核反應(yīng)堆熱工水力專業(yè)