【正文】
如下: y(t)=WNN[y(t1),y(t2),y(t3),u1(t3),u1(t4),u2(t5),u3(t1) (12)其中,y是溫度。 U1是煤氣流量。 U2是加料量。 U3是氧含量。 t為采樣時間。 圖2小波神經(jīng)網(wǎng)絡(luò)模型的結(jié)構(gòu)然后,我們可以推斷出神經(jīng)網(wǎng)絡(luò)的單步預(yù)測模型: ym(t+1)=WNN1[y(t),y(t1),y(t2),u1(t2),u1(t3),u2(t4),u3(t)] (13)多步預(yù)測模型是: ym(t+d)=WNNd[y(t+d1),y(t+d2),y(t+d3),u1(t+d3),u1(t+d4),u2(t+d5),u3(t+d1)] (14)ym(t+1)是時間t的樣本數(shù)據(jù)在時間t+1內(nèi)的預(yù)測結(jié)果, d是預(yù)測步驟。 WNN1是單步預(yù)測模型。 WNNd是d步多步預(yù)測模型。在輸入變量式(14)右側(cè)。 [y,u1,u2,u3,其采樣時間為t + di(i=1,2,3,4,5),如果t + di≤t,其輸入值是真實的樣本值。然而,如果t + di t,其輸入值(t + di),u1(t+di),u2(t + di)和u3(t + di)分別被ym(t+ di),u1,u2(t)(t)和u3(t),取代。因此,時間t內(nèi)的多步預(yù)測模型在單步預(yù)測和多步反復(fù)計算的基礎(chǔ)上可以構(gòu)建。在20世紀(jì)末,近似已經(jīng)形式神經(jīng)網(wǎng)絡(luò)的能力[1921]取得了很大發(fā)展。它已被證明單隱層順向進料神經(jīng)網(wǎng)絡(luò)有近似任何非線性映射功能的特點。因此,一個單隱層神經(jīng)網(wǎng)絡(luò)作為溫度預(yù)測模型在這項工作中獲得通過。作為訓(xùn)練措施,梯度遞減規(guī)律被使用在根據(jù)神經(jīng)網(wǎng)絡(luò)的權(quán)重進行了修改δ規(guī)則的過程中。建模過程包括正向計算和誤差反向傳播。在正向計算時,信息(神經(jīng)元)從輸入層神經(jīng)節(jié)點到輸出節(jié)點通過隱藏的神經(jīng)節(jié)點傳送,每個神經(jīng)元僅影響下一個。如果預(yù)計的輸出層誤差無法獲得,誤差反向傳播將通過神經(jīng)網(wǎng)絡(luò)每個節(jié)點的權(quán)重將被修改。這個過程是重復(fù)的,直到獲得了給定的精度。隱藏節(jié)點數(shù)目用來確定修剪方法[22]。起初,其網(wǎng)絡(luò)隱藏節(jié)點數(shù)量遠(yuǎn)遠(yuǎn)多于其實際要求使用的,然后根據(jù)效能標(biāo)準(zhǔn)方程網(wǎng)絡(luò),沒有或很少貢獻網(wǎng)絡(luò)性能的節(jié)點和權(quán)重被修剪掉,最后可以得到合適的網(wǎng)絡(luò)結(jié)構(gòu)。鑒于BP算法存在的缺陷,如容易下降到一個局部最小值,收斂速度慢,以及低弱的抗干擾能力,采用以下改進措施。1)附加動量項附加動量項應(yīng)用,其功能相當(dāng)于一個低頻率濾波器,認(rèn)為不僅是誤差梯度,而且還改變曲面上誤差傾向,這使得允許改變現(xiàn)有的網(wǎng)絡(luò)。沒有動量功能,網(wǎng)絡(luò)可能會陷入一個局部最小值。使用這種誤差反向傳播方法過程中,改變值與先前權(quán)重的改變成正比的關(guān)系被添加到目前權(quán)重的變化,這被用于一種新的權(quán)重計算。權(quán)重的修改規(guī)則描述如式(15),其中β(0β1=是動量系數(shù)): Δwij(t+1)=wij(t) (15)2)學(xué)習(xí)速率的自適應(yīng)調(diào)整為了改善收斂性能訓(xùn)練過程中的一種自適應(yīng)調(diào)整方法采用學(xué)習(xí)速率。調(diào)整標(biāo)準(zhǔn)是定義如下:當(dāng)某些時候時新的誤差值變的比舊的更大時,學(xué)習(xí)速率將減小,否則,它可保持不變。當(dāng)新的誤差值小于舊的時,學(xué)習(xí)速率將有所增大。這種方法可以使網(wǎng)絡(luò)學(xué)習(xí)保持適當(dāng)?shù)乃俣?。這種策略如式(16)所示。其中指令集是輸出平方誤差在輸出層的總和: η(t+1)=(t) [SSE(t+1)<SSE(t)] η(t+1)=(t) [SSE(t+1)>SSE(t)] (16) η(t+1)=(t) [SSE(t+1)=SSE(t)]為了測試神經(jīng)網(wǎng)絡(luò)模型,使用了450組樣本數(shù)據(jù),其中400組訓(xùn)練50組的預(yù)測。當(dāng)訓(xùn)練循環(huán)次數(shù)達到22375次時,步長可變的訓(xùn)練過程完成后,得到在網(wǎng)絡(luò)學(xué)習(xí)誤差E={7211},即7個輸入層節(jié)點,二十一個隱藏層節(jié)點和輸出層一個節(jié)點。訓(xùn)練有素的網(wǎng)絡(luò)能夠準(zhǔn)確反映焙燒過程并應(yīng)用于預(yù)測。小波神經(jīng)網(wǎng)絡(luò)的預(yù)測結(jié)果如圖3和圖4所示。圖3表示預(yù)測步驟的改變預(yù)測誤差的變化趨勢,從中可以看出,預(yù)測步增加,預(yù)測誤差變大。當(dāng)預(yù)測就步驟低于6時,即在最后就采樣時間后的9分鐘內(nèi),多步預(yù)測平均誤差小于10℃。在圖4所示是一個令人滿意的結(jié)果:作為一個絕對的誤差177?!嬖诤侠矸秶鷥?nèi),單步預(yù)測小波神經(jīng)網(wǎng)絡(luò)就精度可以達到90%。此外,從圖4中可以看出,預(yù)測步驟為6步時準(zhǔn)確性更差,但5步預(yù)測就結(jié)果是可以接受的。有了預(yù)測模型,就可以預(yù)測焙燒溫度的變化趨勢。如果預(yù)測結(jié)果可以顯示溫度高或低,焙燒操作參數(shù)就可以提前調(diào)整,通過這些可以節(jié)約焙燒能源。 圖3多步預(yù)測誤差的變化趨勢圖4小波神經(jīng)網(wǎng)絡(luò)預(yù)測的結(jié)果4結(jié)論1)通過分析樣本數(shù)據(jù),煤氣流量,加料量和氧含量確定為溫度預(yù)測模型主體參數(shù)。模型參數(shù)命令和延遲時間用F檢驗方法推導(dǎo)出。然后采用小波神經(jīng)網(wǎng)絡(luò)確定焙燒過程。實踐中的應(yīng)用表明該模型在焙燒溫度預(yù)測方面表現(xiàn)良好。2)根據(jù)工藝參數(shù)分析,模型具有一定的預(yù)測能力。有了預(yù)測能力,該模型提供了一種系統(tǒng)的分析和優(yōu)化方法,這意味著,當(dāng)影響因素適當(dāng)改變,焙燒溫度的變化趨勢可以分析出來。預(yù)測與基于該模型的分析有指導(dǎo)生產(chǎn)經(jīng)營的意義。參考文獻[1] YANG Chongyu. Process technology of alumina [M]. Beijing:Metallurgy Industry Press, 1994. (in Chinese)[2] ZHANG Liqiang, LI Wenchao. Establishment of some mathematicmodels for Al(OH)3 roasting [J]. Energy Saving of Nonferrous Metallurgy, 1998, 4: 11?15. (in Chinese)[3] WEI Huang. The relations between process parameters, yield and energy consumption in the production of Al(OH)3 [J]. Light Metals,2003(1): 13?18. (in Chinese)[4] TANG Meiqiong, LU Jidong, JIN Gang, HUANG Lai. Software design for Al(OH)3 circulation fluidization roasting system [J].Nonferrous Metals (Extractive Metallurgy), 2004(3): 49?52. (inChinese)[5] WANG Yutao, ZHOU Jianchang, WANG Shi. Application of neural network model and temporal difference method to predict the silicon content of the hot metal [J]. Iron and Steel, 1999, 34(11):7?11. (in Chinese)[6] TU Hai, XU Jianlun, LI Ming. Application of neural network to the forecast of heat state of a blast furnace [J]. Journal of Shanghai University (Natural Science), 1997, 3(6): 623?627. (in Chinese)[7] LU Baiquan, LI Tianduo, LIU Zhaohui. Control based on BPneural networks and wavelets [J]. Journal of System Simulation,1997, 9(1): 40?48. (in Chinese)[8] CHEN Tao, QU Liangsheng. The theory and application of multiresolution wavelet network [J]. China Mechanical Engineering,1997, 8(2): 57?59. (in Chinese)[9] ZHANG Qinghua, BENVENISTE A. Wavelet network [J]. IEEE Transon on Neural Networks, 1992, 3(6): 889?898.[10] PATI Y C, KRISHNA P S. Analysis and synthesis of feed forward network using discrete affine wavelet transformations [J]. IEEE Transon on Neural Networks, 1993, 4(1): 73?85.[11] GROSSMANN A, MORLET J. Deposition of hardy functions into square integrable wavelets of constant shape [J]. SIAM J Math Anal, 1984, 15(4): 723?736.[12] GROUPILLAUD P, GROSSMANN A, MORLET J. Cycleoctave and related transforms in seismic signal analysis [J]. Geoexploration,1984, 23(1): 85?102.[13] GROSSMANN A, MORLET J. Transforms associated to square integrable group representations (I): General results [J]. J Math Phys,1985, 26(10): 2473?2479.[14] ZHAO Songnian, XIONG Xiaoyun. The wavelet transformation and the wavelet analyze [M]. Beijing: Electronics Industry Press,1996. (in Chinese)[15] NIU Dongxiao, XING Mian. A study on wavelet neural network prediction model of time series [J]. Systems Engineering—Theory and Practice, 1999(5): 89?92. (in Chinese)[16] YAO Junfeng, JIANG Jinhong, MEI Chi, PENG Xiaoqi, REN Hongjiu, ZHOU Anliang. Application of wavelet neural network in forecasting slag weight and ponents of coppersmelting converter[J]. Nonferrous Metals, 2001, 53(2): 42?44. (in Chinese)[17] WANG Tianqing. Practice of lowering gaseous suspension calciner heat consumption coast [J]. Energy Saving of Nonferrous Metallurgy, 2004, 21(4): 91?94. (in Chinese)[18] FANG Chongzhi, XIAO Deyun. Process identification [M]. Beijing:Tsinghua University Press, 1988. (in Chinese)[19] CARROLL S M, DICKINSON B W. Construction of neural nets using the radon transform [C]// Proceedings of IJCNN. New York: IEEE Press, 1989: 607?611.[20] ITO Y. Representation of functions by superposition of a step or sigmoidal functions and their applications to neural network theory[J]. Neural Network, 1991, 4: 385?394.[21] JAROSLAW P S, KRZYSZTOF J C. On the synthesis and plexity of feedforward networks [C]// IEEE World Congress on Computational Intelligence. IEEE Neural Network, 1994:2185?2190.[22] HAYKIN S. Neural networks: A prehensive foundation [M]. 2nd Edition. Beijing: China Machine Press, 200419