張 迪,郭 婷,鄭 萍,姜佰文
隨機(jī)模型預(yù)測(cè)UASB反應(yīng)器對(duì)奶牛養(yǎng)殖廢水處理效果
張 迪1,郭 婷1,鄭 萍2,姜佰文1
(1.東北農(nóng)業(yè)大學(xué)資源與環(huán)境學(xué)院,哈爾濱 150030;2.東北農(nóng)業(yè)大學(xué)電氣與信息化學(xué)院,哈爾濱 150030)
為實(shí)現(xiàn)UASB反應(yīng)器運(yùn)行人工智能控制,采用三層BP神經(jīng)網(wǎng)絡(luò)(Back Propagation Artificial Neural Network,BP-ANN)預(yù)測(cè)升流厭氧反應(yīng)器處理奶牛養(yǎng)殖廢水COD去除效果。運(yùn)用BP神經(jīng)網(wǎng)絡(luò)構(gòu)建出水與進(jìn)水COD濃度、水力停留時(shí)間、pH、溫度、堿度、揮發(fā)性有機(jī)酸、有機(jī)負(fù)載率和懸浮固體之間非線性模型,比較不同算法。Levenberg-Marquardt算法為BP神經(jīng)網(wǎng)絡(luò)最佳算法,最佳結(jié)構(gòu)為8-8-1,模擬訓(xùn)練效果較好。BP神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)值與真實(shí)值接近,一致性較高,模型擬合程度較好。利用線性-非線性模型評(píng)價(jià)不同輸入?yún)?shù)對(duì)廢水COD去除率影響,比較BP-ANN與線性-非線性模型預(yù)測(cè)效果,為奶牛養(yǎng)殖廢水處理智能化管理提供技術(shù)支持。
人工神經(jīng)網(wǎng)絡(luò);UASB反應(yīng)器;COD去除率;Levenberg-Marquardt算法;奶牛養(yǎng)殖廢水
目前畜禽養(yǎng)殖廢水已成為農(nóng)村面源污染主要來(lái)源之一[1],該類廢水一般經(jīng)固液分離,進(jìn)入?yún)捬跆幚砉に噯卧?缺氧處理-好氧處理-沉淀池后排出中水供農(nóng)牧業(yè)回用[2-3]。隨著《中華人民共和國(guó)水污染防治法》及畜禽養(yǎng)殖業(yè)污染物排放標(biāo)準(zhǔn)修訂,奶牛養(yǎng)殖廢水排放要求日益嚴(yán)格,養(yǎng)殖企業(yè)急需解決廢水達(dá)標(biāo)排放問(wèn)題。廢水處理系統(tǒng)具有復(fù)雜性、非線性、時(shí)變性、不確定性和滯后性等特點(diǎn),難以有效控制全部處理過(guò)程。常規(guī)數(shù)學(xué)模型法無(wú)法獲得精確數(shù)學(xué)模型[4-5]。人工神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)方法具有自學(xué)習(xí)、自適應(yīng)和自組織功能,適合復(fù)雜非線性系統(tǒng)建模和控制[6],已成為廢水處理過(guò)程控制研究熱點(diǎn)。
針對(duì)靜態(tài)吸附試驗(yàn)過(guò)程控制,Esfandian等利用人工神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)改性沸石分子篩對(duì)水相介質(zhì)中二嗪農(nóng)殺蟲(chóng)劑去除效果[7]。Yetilmezsoy等建立開(kāi)心果殼對(duì)廢水中鉛吸附去除效果人工神經(jīng)網(wǎng)絡(luò)模型[8]。在廢水動(dòng)態(tài)反應(yīng)器處理過(guò)程控制方面,Shi等考查UASBAF反應(yīng)器對(duì)含高濃度中藥廢水處理過(guò)程中不同參數(shù)對(duì)BP神經(jīng)網(wǎng)絡(luò)模型適應(yīng)性學(xué)習(xí)速度和動(dòng)量影響[9]。易賽莉等利用人工神經(jīng)網(wǎng)絡(luò)構(gòu)建UASB反應(yīng)器處理生活污水隨機(jī)模型[10],在仿真條件下較好預(yù)測(cè)污染物去除率。綜上所述,人工神經(jīng)網(wǎng)絡(luò)具有大量人工神經(jīng)元,廣泛相互連接形成復(fù)雜網(wǎng)絡(luò),為廢水處理系統(tǒng)多輸入、多輸出系統(tǒng)提供有效解決方案,適合復(fù)雜系統(tǒng)優(yōu)化及內(nèi)部未知系統(tǒng)逼近和模擬[11-13]。
因此,本研究利用BP人工神經(jīng)網(wǎng)絡(luò)對(duì)奶牛養(yǎng)殖廢水UASB反應(yīng)器進(jìn)出水中COD等相關(guān)指標(biāo)作樣本訓(xùn)練和測(cè)試,訓(xùn)練適合該污水處理工藝神經(jīng)網(wǎng)絡(luò)模型結(jié)構(gòu),模擬和預(yù)測(cè)影響因素多、機(jī)理復(fù)雜,高難度非線性污水系統(tǒng),選擇容錯(cuò)性、自適性、并行性較強(qiáng)軟測(cè)量技術(shù),為奶牛養(yǎng)殖廢水處理智能化管理提供技術(shù)保障。
BP(Back Propagation)網(wǎng)絡(luò)是1986年Rumelhart和McCelland等提出,按誤差逆?zhèn)鞑ニ惴ㄓ?xùn)練多層前饋網(wǎng)絡(luò),是目前應(yīng)用最廣泛神經(jīng)網(wǎng)絡(luò)模型之一。BP網(wǎng)絡(luò)可學(xué)習(xí)和存貯輸入-輸出模式映射關(guān)系,不必事前揭示描述這種映射關(guān)系數(shù)學(xué)方程,其學(xué)習(xí)規(guī)則運(yùn)用最速下降法,通過(guò)反向傳播調(diào)整網(wǎng)絡(luò)權(quán)值和閾值,網(wǎng)絡(luò)誤差平方和最小。BP神經(jīng)網(wǎng)絡(luò)模型拓?fù)浣Y(jié)構(gòu)包括輸入層(input)、隱層(hide layer)和輸出層(output layer)。BP算法包含數(shù)據(jù)流前向計(jì)算(正向傳播)和誤差信號(hào)反向傳播兩個(gè)過(guò)程。正向傳播時(shí),傳播方向?yàn)檩斎雽印[層→輸出層,每層神經(jīng)元狀態(tài)影響下一層神經(jīng)元。若在輸出層得不到期望輸出,則轉(zhuǎn)向誤差信號(hào)反向傳播流程。通過(guò)兩過(guò)程交替作用,在權(quán)向量空間執(zhí)行誤差函數(shù)梯度下降策略,動(dòng)態(tài)迭代搜索一組權(quán)向量,使網(wǎng)絡(luò)誤差函數(shù)達(dá)最小值,完成信息提取和記憶過(guò)程。即輸入信號(hào)Xi通過(guò)中間節(jié)點(diǎn)(隱層點(diǎn))作用于輸出節(jié)點(diǎn),經(jīng)過(guò)非線形變換,產(chǎn)生輸出信號(hào)Yk,網(wǎng)絡(luò)訓(xùn)練每個(gè)樣本包括輸入向量X和期望輸出量t,網(wǎng)絡(luò)輸出值Y與期望輸出值t之間偏差,通過(guò)調(diào)整輸入節(jié)點(diǎn)與隱層節(jié)點(diǎn)聯(lián)接強(qiáng)度取值Wij和隱層節(jié)點(diǎn)與輸出節(jié)點(diǎn)間聯(lián)接強(qiáng)度Tjk及閾值,使誤差沿梯度方向下降,經(jīng)反復(fù)學(xué)習(xí)訓(xùn)練,確定與最小誤差對(duì)應(yīng)網(wǎng)絡(luò)參數(shù)(權(quán)值和閾值),訓(xùn)練即告停止。此時(shí)經(jīng)訓(xùn)練神經(jīng)網(wǎng)絡(luò)可對(duì)類似樣本輸入信息,自動(dòng)處理輸出誤差最小非線形轉(zhuǎn)換信息。
BP網(wǎng)絡(luò)模型包括輸入輸出模型、作用函數(shù)模型、誤差計(jì)算模型和自學(xué)習(xí)模型。
①節(jié)點(diǎn)輸出模型
X-輸入量;O-輸出量;Y-輸出層輸出量;f-非線形作用函數(shù);q-神經(jīng)單元閾值。
②作用函數(shù)模型
作用函數(shù)是反映下層輸入對(duì)上層節(jié)點(diǎn)刺激脈沖強(qiáng)度函數(shù),又稱刺激函數(shù),一般取為(0,1)內(nèi)連續(xù)取值Sigmoid函數(shù):
③誤差計(jì)算模型是反映神經(jīng)網(wǎng)絡(luò)期望輸出與計(jì)算輸出之間誤差函數(shù):
tpi-i節(jié)點(diǎn)期望輸出值;Opi-i節(jié)點(diǎn)計(jì)算輸出值。
④自學(xué)習(xí)模型神經(jīng)網(wǎng)絡(luò)學(xué)習(xí)過(guò)程,即連接下層節(jié)點(diǎn)和上層節(jié)點(diǎn)間權(quán)重拒陣Wij設(shè)定和誤差修正過(guò)程。BP網(wǎng)絡(luò)分有師學(xué)習(xí)方式-需要設(shè)定期望值和無(wú)師學(xué)習(xí)方式-僅輸入模式。
自學(xué)習(xí)模型為:
h-學(xué)習(xí)因子;φi-輸出節(jié)點(diǎn)i計(jì)算誤差;Oj-輸出節(jié)點(diǎn)j計(jì)算輸出;a-動(dòng)量因子。
BP網(wǎng)絡(luò)模型應(yīng)用存在數(shù)據(jù)量過(guò)大、節(jié)點(diǎn)數(shù)過(guò)多網(wǎng)絡(luò)學(xué)習(xí)時(shí)間過(guò)長(zhǎng),甚至不能收斂,對(duì)運(yùn)算速度和數(shù)據(jù)存儲(chǔ)要求較高,容錯(cuò)能力相對(duì)較差等問(wèn)題。因此,在應(yīng)用中需學(xué)習(xí)因子h優(yōu)化、隱含層節(jié)點(diǎn)數(shù)優(yōu)化、確定輸入和輸出神經(jīng)元。本研究主要解決BP-ANN神經(jīng)網(wǎng)絡(luò)在UASB反應(yīng)器處理奶牛養(yǎng)殖廢水COD去除率軟測(cè)量時(shí)模型節(jié)點(diǎn)優(yōu)化,輸入神經(jīng)元篩選以及算法優(yōu)化等問(wèn)題。
構(gòu)建UASB反應(yīng)器奶牛養(yǎng)殖廢水COD去除效果模型首先需確定影響處理效果模型參數(shù)。UASB(升流式厭氧污泥床反應(yīng)器)是目前應(yīng)用最廣泛處理高濃度有機(jī)廢水高速厭氧反應(yīng)器[14-16]。影響UASB反應(yīng)器性能主要因素有溫度、pH、營(yíng)養(yǎng)物與微量元素、堿度和揮發(fā)性酸濃度、進(jìn)水中懸浮固體濃度、氨氮濃度、硫酸鹽濃度、其他有毒物質(zhì)等[17-19]。本研究采用數(shù)據(jù)均來(lái)自UASB反應(yīng)器,反應(yīng)器運(yùn)行時(shí)間2014年3月~2015年3月,考慮不同季節(jié)變化對(duì)反應(yīng)器運(yùn)行影響,數(shù)據(jù)選擇包含春夏秋冬4季。最終選取162組數(shù)據(jù),其中訓(xùn)練數(shù)據(jù)120組,仿真數(shù)據(jù)42組,數(shù)據(jù)統(tǒng)計(jì)如表1所示。
表1 UASB反應(yīng)器運(yùn)行參數(shù)統(tǒng)計(jì)分析Table 1 Running data statistics of UASB reactor
篩選數(shù)據(jù)參數(shù)屬性不同,量綱不同,數(shù)據(jù)需作歸一化處理。為便于模型運(yùn)行和計(jì)算,BP神經(jīng)網(wǎng)絡(luò)要求傳遞函數(shù)可微分,一般為S型函數(shù),即logsig函數(shù)。在人工神經(jīng)網(wǎng)絡(luò)中對(duì)數(shù)據(jù)作歸一化處理,使其輸入值一般在區(qū)間[0,1]或[-1,1]之間,提高神經(jīng)網(wǎng)絡(luò)訓(xùn)練收斂速度。本研究采用min-max規(guī)范化[20]。歸一化方法如下:
取值范圍[0,1],其中,xi為觀察值,xmin為最小觀察值,xmax為最大觀察值。由于數(shù)據(jù)保留位數(shù)差異,出現(xiàn)極接近0或1值,修約為0或1即可。
采用MATLAB 2012中Neural Network Toolbox V4.0預(yù)測(cè)COD去除率。在實(shí)驗(yàn)室條件下,利用UASB反應(yīng)器處理真實(shí)奶牛養(yǎng)殖廢水365 d,共獲得162組數(shù)據(jù)。模型參數(shù)確定采用Matlab中包含特征降維(PCA)算法實(shí)現(xiàn),主要通過(guò)調(diào)用princomp函數(shù)實(shí)現(xiàn)。篩選進(jìn)水COD、水力停留時(shí)間(HRT)、堿度、pH、水力負(fù)荷量、揮發(fā)性有機(jī)酸、懸浮固體、溫度共計(jì)8個(gè)變量參數(shù)作為輸入矩陣[p],COD去除率作為輸出矩陣[t]。采用BP神經(jīng)網(wǎng)絡(luò)是多層前饋神經(jīng)網(wǎng)絡(luò)模型,由一個(gè)輸入層、一個(gè)輸出層和一個(gè)或多個(gè)隱含層構(gòu)成。BP神經(jīng)網(wǎng)絡(luò)層數(shù)由隱含層層數(shù)確定,隱含層為1時(shí),BP神經(jīng)網(wǎng)絡(luò)精確度較高,隱含層數(shù)量過(guò)多則BP神經(jīng)網(wǎng)絡(luò)學(xué)習(xí)時(shí)間增加,精確度降低。故本研究采用三層BP神經(jīng)網(wǎng)絡(luò),即輸入層、隱含層和輸出層。BP神經(jīng)網(wǎng)絡(luò)由處理問(wèn)題復(fù)雜程度和學(xué)習(xí)數(shù)目決定[21]。本研究采用均方差作為評(píng)價(jià)標(biāo)準(zhǔn),評(píng)價(jià)Rprop、Fletcher-Reeves、Polak-Ribiere、Powell-Beale、Levenberg-Marquardt等算法,篩選最適BP神經(jīng)網(wǎng)絡(luò)訓(xùn)練算法。
如表2所示,比較5種改進(jìn)算法,以便篩選最適BP神經(jīng)網(wǎng)絡(luò)算法。所有訓(xùn)練算法均采用隱含層正切S型函數(shù)(tansig)且輸出層為線性傳遞函數(shù)(purelin)三層網(wǎng)絡(luò)結(jié)構(gòu)。所有算法隱含層均為10個(gè)神經(jīng)元。五種不同算法評(píng)價(jià)參數(shù)見(jiàn)表2和圖1。比較以最小均方差作為評(píng)判算法優(yōu)劣標(biāo)準(zhǔn)。LMA算法因均方差最小被篩選為最優(yōu)算法,5種算法均方差比較見(jiàn)圖1。本研究比較表2所列不同算法,Levenberg-Marquardt算法具有最小平均誤差,因此LM為最佳算法。
表2 BP神經(jīng)網(wǎng)絡(luò)模型算法比較Table 2 Comparisons of five BP algorithms with various neurons in the hidden layer
圖1 BP神經(jīng)網(wǎng)絡(luò)算法比較Fig.1 Comparison between BP algorithms
本研究使用6個(gè)神經(jīng)元作為隱含層,將神經(jīng)元數(shù)目逐漸增至20,獲得不同均方誤差如圖2。隨神經(jīng)元數(shù)目變化,均方誤差不斷改變。LM算法中,當(dāng)隱含層神經(jīng)元數(shù)目為8時(shí),均方誤差最小,為0.005,因此隱含層神經(jīng)元數(shù)目設(shè)定為8。由此獲得最佳人工神經(jīng)網(wǎng)絡(luò)模型結(jié)構(gòu)流程,如圖3所示。三層網(wǎng)絡(luò)結(jié)構(gòu),隱含層為神經(jīng)元為8正切S型傳遞函數(shù),輸出層線性傳遞函數(shù)。精確數(shù)學(xué)表達(dá)式如公式(1)和(2)所示。線性傳遞函數(shù)被用于輸出層,但沒(méi)有負(fù)ANN預(yù)測(cè)出現(xiàn)在ANN輸出值,可歸因于本研究輸出向量特性。
如圖3所示。三層網(wǎng)絡(luò)結(jié)構(gòu),隱含層為神經(jīng)元為8的sigmoid傳遞函數(shù),輸出層線性傳遞函數(shù)。精確數(shù)學(xué)表達(dá)式如公式(7)和(8)所示。
圖2 LM算法中隱含層神經(jīng)元數(shù)目確定Fig.2 Number of neurons at hidden layer for the Levenberg-Marquardt algorithm
圖3 最佳人工神經(jīng)網(wǎng)絡(luò)模型結(jié)構(gòu)流程Fig.3 Optimal BP model structure for the prediction of CODRE
BP神經(jīng)網(wǎng)絡(luò)經(jīng)過(guò)學(xué)習(xí)訓(xùn)練后,測(cè)試與驗(yàn)證訓(xùn)練成果。測(cè)試數(shù)據(jù)為22組,測(cè)試后得預(yù)測(cè)值,預(yù)測(cè)值與真實(shí)數(shù)據(jù)比較結(jié)果見(jiàn)圖4。
由圖4可知,BP神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)值接近實(shí)驗(yàn)所測(cè)真實(shí)值,變化趨勢(shì)與真實(shí)值基本一致,相關(guān)系數(shù)較大,最小相關(guān)系數(shù)為0.694。因此,BP神經(jīng)網(wǎng)絡(luò)模型符合奶牛場(chǎng)污水處理特點(diǎn)。
圖4 BP神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)值與真實(shí)值比較Fig.4 Agreements between ANN testing outputs and the experimental data
利用DataFit 9.0數(shù)據(jù)分析軟件(Version 9.0.59,1998-2008 Oakdale Engineering)對(duì)UASB反應(yīng)器運(yùn)行參數(shù)和COD去除率結(jié)果回歸擬合。比較實(shí)驗(yàn)輸出值和神經(jīng)網(wǎng)絡(luò)擬合值如圖5所示。Levenberg-Marquardt方法精確度是線性和非線性方法2倍。根據(jù)擬合優(yōu)度自動(dòng)排序,2個(gè)線性模型和1個(gè)指數(shù)模型作回歸分析,結(jié)果參見(jiàn)表3。
表3 回歸分析結(jié)果Table 3 Summary of regression results
根據(jù)相關(guān)系數(shù)判定模型1是最佳擬合模型(R2=0.48),UASB反應(yīng)器運(yùn)行162組數(shù)據(jù)與回歸模型1輸出值比較如圖5所示。模型相關(guān)系數(shù)和統(tǒng)計(jì)量結(jié)果參見(jiàn)表3和表4。T-比率代表參與預(yù)測(cè)參數(shù)變量與預(yù)測(cè)參數(shù)標(biāo)準(zhǔn)變異系數(shù)比值。由表4可知,影響COD去除率顯著性參數(shù)分別為水力負(fù)荷、溫度、水力停留時(shí)間。P值越小表征越顯著相關(guān),因此,入水COD、水力停留時(shí)間(HRT)、溫度、水力負(fù)荷均為極顯著相關(guān)。模型T值和P值顯示CODi濃度,水力停留時(shí)間、運(yùn)行溫度、水力負(fù)荷對(duì)系統(tǒng)運(yùn)行重要性大于OLR和pH。
圖5顯示162個(gè)觀測(cè)數(shù)據(jù)與回歸模型輸出數(shù)據(jù)一致性問(wèn)題,回歸模型輸出值顯示,COD去除率與實(shí)際實(shí)驗(yàn)數(shù)據(jù)之間方差最大為0.3678,相關(guān)系數(shù)(R2=0.541)。
隨機(jī)模型研究結(jié)果顯示,與線性和非線性模型相比,ANN模型變異系數(shù)較小,更適于預(yù)測(cè)UASB反應(yīng)器對(duì)奶牛養(yǎng)殖廢水處理COD去除率。
表4 最適模型參數(shù)及回歸變量Table 4 Model coefficients and regression statistical results for the best model
圖5 回歸模型預(yù)測(cè)值與試驗(yàn)數(shù)據(jù)比較(162個(gè)觀察值)Fig.5 Agreement between the regression model outputs and the 162 experimental data
人工神經(jīng)網(wǎng)絡(luò)模型便于修改和補(bǔ)充。ANN模型因其開(kāi)發(fā)過(guò)程中參數(shù)較少,與白箱模型相比更簡(jiǎn)便。當(dāng)測(cè)量變量與ANN預(yù)測(cè)響應(yīng)值不同時(shí),模型可使用交叉數(shù)據(jù)對(duì)模型重新訓(xùn)練。這個(gè)過(guò)程通過(guò)嵌入ANN模型到專家系統(tǒng),可實(shí)現(xiàn)整個(gè)過(guò)程完整控制。因此本研究選用隨機(jī)模型BP-ANN人工神經(jīng)網(wǎng)絡(luò)模型預(yù)測(cè)UASB反應(yīng)器對(duì)奶牛養(yǎng)殖廢水處理效果。
選擇適合算法獲得最快運(yùn)算速度是研究既定問(wèn)題難點(diǎn)。影響因素涉及問(wèn)題復(fù)雜性及選擇訓(xùn)練數(shù)據(jù)量。標(biāo)準(zhǔn)BP算法基于梯度下降法,通過(guò)計(jì)算目標(biāo)函數(shù)修正網(wǎng)絡(luò)權(quán)值和閾值梯度。但存在易形成局部極小,訓(xùn)練陷入癱瘓問(wèn)題。因此,基于標(biāo)準(zhǔn)值優(yōu)化改進(jìn)算法和基于標(biāo)準(zhǔn)梯度下降改進(jìn)方法應(yīng)用廣泛[22]。
在共軛梯度算法中,Powell-Beale程序需要存儲(chǔ)最多,但通常收斂速度最快。Polak-Ribiere和Rprop有相似表現(xiàn),但這類算法難以應(yīng)用于既定問(wèn)題預(yù)測(cè)。Polak-Ribiere算法相比較于Fletcher-Reeves有更大存儲(chǔ)需求。Fletcher-Reeves收斂次數(shù)比Rprop算法少,Rprop算法每次迭代所需計(jì)算量大。Rprop算法和尺度共軛梯度算法不需要一條直線搜索,并有最小存儲(chǔ)需求。理論上運(yùn)算速度更快,數(shù)據(jù)量大模擬更適合??勺儗W(xué)習(xí)速率算法較慢,但更適用于處理某些問(wèn)題。Levenberg-Marquardt算法因包含數(shù)百個(gè)權(quán)重值而具有最快的收斂速度,本研究需要精確訓(xùn)練,故Levenberg-Marquardt算法具有最大優(yōu)勢(shì)。
Almasri等發(fā)現(xiàn)人工神經(jīng)網(wǎng)絡(luò)核心是人工網(wǎng)絡(luò)學(xué)習(xí)部分,決定最佳模型結(jié)構(gòu)。因此,隱含層中神經(jīng)元數(shù)目設(shè)定尤為重要[23],在系統(tǒng)運(yùn)行中起重要作用,故作神經(jīng)元數(shù)目和最小均方差優(yōu)化。此外,隱含層和輸出層函數(shù)選擇對(duì)BP神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)精度影響較大。本研究中盡管輸出層采用線性傳遞函數(shù),而線性輸出層加上線性輸出函數(shù)易使輸出值范圍超過(guò)[-1,+1]。但對(duì)于已構(gòu)建人工神經(jīng)網(wǎng)絡(luò)來(lái)說(shuō),未出現(xiàn)負(fù)預(yù)測(cè)值,預(yù)測(cè)精度較高,可能是輸入向量特性決定。高均方差易出現(xiàn)以正切函數(shù)為傳遞函數(shù)輸出層人工神經(jīng)網(wǎng)絡(luò)模型[24-26]。本研究中選擇較小均方差作為衡量構(gòu)建人工神經(jīng)網(wǎng)絡(luò)模型優(yōu)劣依據(jù),故選擇輸出層函數(shù)為purelin線性傳遞函數(shù)。線性輸出層與線性函數(shù)purelin使產(chǎn)值在[-1,+1],故輸入向量特征決定負(fù)ANN估計(jì)預(yù)期。另一方面,假如約束網(wǎng)絡(luò)輸出值是理想值(比如在0和1之間),輸出層應(yīng)該采用Sigmoid傳遞函數(shù)(如logsig),輸出神經(jīng)元線性激活函數(shù)被替換。在本研究具體問(wèn)題中,COD去除率作為輸出值始終保持正值,故該輸出神經(jīng)元線性激活函數(shù)被替換。此外,最優(yōu)人工神經(jīng)網(wǎng)絡(luò)模型體系結(jié)構(gòu)及其參數(shù)變化確定基于MSE最小值,第二網(wǎng)絡(luò)結(jié)構(gòu)可取,即使負(fù)估計(jì)值不符合實(shí)際輸出總是大于0,處理方式也可應(yīng)用于計(jì)算研究[27-29]。
回歸模型預(yù)測(cè)值與實(shí)際測(cè)量值之間殘差為-0.07~0.06,相關(guān)系數(shù)0.481。利用BP神經(jīng)網(wǎng)絡(luò)擬合值與測(cè)量值之間殘差為-0.04~0.09,相關(guān)系數(shù)為0.694。因此BP神經(jīng)網(wǎng)絡(luò)更為適合作為UASB反應(yīng)器處理高濃度有機(jī)廢水COD去除率軟測(cè)量模型。未來(lái)研究則需提高數(shù)據(jù)廣泛性及準(zhǔn)確性,獲得足夠訓(xùn)練樣本并選擇最適訓(xùn)練樣本。
[1] 王麗麗,梁俊爽,王忠江.黑龍江厭氧技術(shù)處理畜禽糞便現(xiàn)狀及發(fā)展方向探討[J].東北農(nóng)業(yè)大學(xué)學(xué)報(bào),2006,37(2):283-288.
[2] 劉海慶,李文哲,王忠江.中、高溫酸化兩相厭氧發(fā)酵處理牛糞的試驗(yàn)[J].東北農(nóng)業(yè)大學(xué)學(xué)報(bào),2007,38(6):809-813.
[3] 謝秋菊,羅文博,李妍.基于神經(jīng)網(wǎng)絡(luò)豬舍氨氣濃度預(yù)測(cè)方法研究[J].東北農(nóng)業(yè)大學(xué)學(xué)報(bào),2016,47(10):83-92.
[4] Giwa A,Daer S,Ahmed I,et al.Experimental investigation and artificial neural networks ANNs modeling of electrically-enhanced membrane bioreactor for wastewater treatment[J].Journal of Water Process Engineering,2016,11:88-97.
[5] Hong G H,Ying L,Jun F Q,A fuzzy neural network approach for online fault detection in waste water treatment process[J].Computers&Electrical Engineering,2014,40(7):2216-2226.
[6] Turan N G,Mesci B,Ozgonenel O.The use of artificial neural networks(ANN)for modeling of adsorption of Cu(II)from industrial leachate by pumice[J].Chemical Engineering Journal,2011,171(3):1091-1097.
[7] Esfandian H,Samadi-Maybodi A,Parvini M,et al.Development of a novel method for the removal of diazinon pesticide from aqueous solution and modeling by artificial neural networks(ANN)[J].Journal of Industrial and Engineering Chemistry,2016,35:295-308.
[8] Yetilmezsoy K,Demirel S.Artificial neural network(ANN)approach for modeling of Pb(II)adsorption from aqueous solution by Antep pistachio(Pistacia Vera L.)shells[J].Journal of Hazardous Materials,2008,153(3):1288-1300.
[9] Shi Y,Zhao X T,Zhang Y M,et al.Back propagation neural network(BPNN)prediction model and control strategies of methanogen phase reactor treating traditional Chinese medicine wastewater(TCMW)[J].Journal of Biotechnology,2009,144(1):70-74.
[10] 易賽莉,雒文生.人工神經(jīng)網(wǎng)絡(luò)在UASB反應(yīng)器處理生活污水中的模擬預(yù)測(cè)與應(yīng)用[J].武漢大學(xué)學(xué)報(bào):工學(xué)版,2004,37(1):46-50.
[11] Howard H.Chapter 7-ANNs for quantifier learning and recognition[M].In Neuromimetic Semantics,Elsevier Science B V:Amsterdam,2004.
[12] Mittal G S.Chapter 18-Artificial Neural Network(ANN)Based Process Modeling A2-Kutz,Myer.In Handbook of Farm,Dairy and Food Machinery Engineering(Second Edition)[M].San Diego:Academic Press,2013.
[13] Noori N,Kalin L.Coupling SWAT and ANN models for enhanced daily stream flow prediction[J].Journal of Hydrology,2016,533:141-151.
[14] Al-Shayah M,Mahmoud N.Start-up of an UASB-septic tank for community on-site treatment of strong domestic sewage[J].Bioresource Technology,2008,99(16):7758-7766.
[15] Alvarez J A,Armstrong E,Presas J,et al.Performance of a UASBDigester system treating domestic wastewater[J].Environmental Technology,2004,25(10):1189-1199.
[16] Fu Q,Fu H,Liang C.Considering the influence of Multiweather-factors to forecasting the water requirement of well irrigation rice based on ANN model[J].Journal of Northeast Agricultural University:English Edition,2004,11(1):57-61.
[17] Zekker I,Rikmann E,Tenno T,et al.Start-up of low-temperature anammox in UASB from mesophilic yeast factory anaerobic tank inoculum[J].Environmental Technology,2015,36(2):214-225.
[18] Pontes P P,Chernicharo C A L,Von Sperling M.Kinetics of concentration decay of specific organic matter in UASB reactors operating with and without return of aerobic sludge[J].Environmental Technology,2014,35(16):2046-2054.
[19] Zhang Y,Wang X,Hu M,et al.Effect of hydraulic retention time(HRT)on the biodegradation of trichloroethylene wastewater and anaerobic bacterial community in the UASB reactor[J].Applied Microbiology Biotechnology,2015,99(4):1977-1987.
[20] Peng Z Q,Cao C,Huang J Y et al.Seismic signal recognition using improved BP neural network and combined feature extraction method[J].Journal Central South University of Technology,2014,21(5):1898-1906.
[21] Livingstone D.Artificial neural networks:Methods and applications[M].Totowa:Humana Press,2008.
[22] Tosh C,Ruxton G D.Modelling perception with artificial neural networks[M].New York:Cambridge University Press,2010.
[23] Almasri M N,Kaluarachchi J J.Multi-criteria decision analysis for the optimal management of nitrate contamination of aquifers[J].Journal Environmental Management,2005,74(4):365-381.
[24] 付強(qiáng),劉建禹,王立昆,等.基于人工神經(jīng)網(wǎng)絡(luò)的井灌水稻區(qū)地下水位預(yù)測(cè)[J].東北農(nóng)業(yè)大學(xué)學(xué)報(bào),2002,33(2):152-159.
[25] Hu L,Wang F,Xu J,et al.Prediction of liver injury using the BP-ANN model with metabolic parameters in overweight and obese Chinese subjects[J].International Journal of Clinical and Experimental Medicinei,2015,8(8):13359-13364.
[26] Cheng J,Qi B,Chen D,et al.Modification of an RBF ANN-Based temperature compensation model of interferometric fiber optical gyroscopes[J].Sensors(Basel),2015,15(5):11189-111207.
[27] Nagy G,Kutor L.Multiple ANN recognizers for adaptive recognition of thespeech of dysarthric patients in AAL systems[J].Studies in Health Technology Informatics,2015,217:1013-1016.
[28] Maghsoudi M,Ghaedi M,Zinali A,et al.Artificial neural network(ANN)method for modeling of sunset yellow dye adsorption using zinc oxide nanorods loaded on activated carbon:Kinetic and isotherm study[J].Spectrochimica Acta Part A:Molecular and Biomolecular Spectroscopy,2015,134:1-9.
[29] Kourgialas N N,Dokou Z,Karatzas G P.Statistical analysis and ANN modeling for predicting hydrological extremes under climate change scenarios:The example of a small Mediterranean agrowatershed[J].Journal of Environmental Management,2015,154:86-101.
Stochastic modeling applications for prediction of chemical oxygen demand removal efficiency of UASB reactors treating real Dairy wastewater/
ZHANG Di1,GUO Ting1,ZHENG Ping2,JIANG Baiwen1
(1.School of Resources and Environment,Northeast Agricultural University,Harbin 150030,China;2.School of Electrical Engineering and Information,NortheastAgricultural University,Harbin 150030,China)
Wastewater effluents from intensive dairy farm are characterized by high volumes and extremely variable composition.The discharge of dairy farm wastewater into the environment damages the quality of the receiving water and might be toxic to food chain organism and aquatic life.Therefore upflow Anaerobic Sludge Blanket(UASB)reactors are widely used to treat this kind of wastewater.A three-layer Artificial Neural Network model for the prediction of Chemical Oxygen Demand Removal Efficiency(CODRE)of UASB reactors treating real dairy wastewater was presented in this paper.To validate the proposed method,an experimental study was carried out in lab-scale UASB reactors to investigate the efficiency of the total COD reduction.The reactors were operated for 365 at the mesophilic conditions and 162 sets of data could be chosen to build the model.CODRE of UASB reactors being output parameter of the conducted anaerobic treatment was estimated by eight input parameters such as HRT,pH,operating temperatureand so on,according to Backpropagation(BP)training combined with principal component analysis(PCA).In the Artificial Neural Network(ANN)study,theLevenberg-Marquardt Algorithm(LMA)was found as the best of five BP(Back Propagation)algorithms,and the best structure was 8-8-1.ANN model predicted CODRE values based on experimental data and all the predictions were proven to be satisfactory.So BP-ANN model in this paper could be used to prediction the CODRE values in the future real factories.In addition to determination of the optimal ANN structure,a linear-nonlinear study was also employed to investigate the effects of input variables on CODRE values in this study.Both ANN outputs and linear-nonlinear study results were compared and advantages and further developments were evaluated in this paper.
Back Propagation Artificial Neural Network(BP-ANN);Upflow Anaerobic Sludge Blanket(UASB);Chemical Oxygen Demand Reduction(CODRE);Levenberg-Marquardt Algorithm;dairy farm wastewater
S879.3
A
1005-9369(2017)11-0043-09
時(shí)間2017-12-7 12:36:26 [URL]http://kns.cnki.net/kcms/detail/23.1391.S.20171207.1236.010.html
張迪,郭婷,鄭萍,等.隨機(jī)模型預(yù)測(cè)UASB反應(yīng)器對(duì)奶牛養(yǎng)殖廢水處理效果[J].東北農(nóng)業(yè)大學(xué)學(xué)報(bào),2017,48(11):43-51.
Zhang Di,Guo Ting,Zheng Ping,et al.Stochastic modeling applications for prediction of chemical oxygen demand removal efficiency of UASB reactors treating real Dairy wastewater[J].Journal of Northeast Agricultural University,2017,48(11):43-51.(in Chinese with English abstract)
2017-09-15
國(guó)家重點(diǎn)研發(fā)計(jì)劃項(xiàng)目(2017YFD0300503);國(guó)家科技支撐項(xiàng)目(2013BAD21B01)
張迪(1979-),女,副教授,博士研究生,研究方向?yàn)檗r(nóng)業(yè)環(huán)境污染物遷移轉(zhuǎn)化。E-mail:zhangdi6283@neau.edu.cn