(中南工業(yè)大學 應用物理與熱能工程系， 長沙 410083)
摘 要: 討論了權(quán)值初始化、 變量的預處理、學習過程參數(shù)的自適應調(diào)節(jié)、 網(wǎng)絡(luò)拓撲結(jié)構(gòu)等因素對學習和推廣的影響， 提出了一種改進的BP神經(jīng)網(wǎng)絡(luò)學習算法， 在很大程度上改善了學習效率。采用改進的帶有8個輸入變量的BP神經(jīng)網(wǎng)絡(luò)算法和自適應殘差補償算法建立吹煉終點組合預報模型。利用某廠實際生產(chǎn)數(shù)據(jù)進行仿真運行的結(jié)果表明， 利用該組合預報模型得到的平均相對預測誤差為1.2%， 最大誤差為4%。
關(guān)鍵字: 轉(zhuǎn)爐； 銅锍吹煉； 神經(jīng)網(wǎng)絡(luò)；終點預報
（Department of Applied Physics and Heat Engineering， Central South University of Technology， Changsha 410083， P.R.China）
Abstract:The effect of mass initialization, variables pretreatment, adaptive adjustment of parameters and structure of network on exercvse and generalization was discussed in detail. An improved BP neural network exercise algorithm which was developed to greatly improve its efficiency was proposed, and a grouping prediction model based on the neural network algorithm with 8 input variables and error compensation of linear regression were developed. For simulating test, the average error of the model is 1.2%, and the maximal error of the model is 4%.
Key words: converting furnace； neural network； endpointprediction； linear regression
