基于隐Markov模型的高炉铁水硅质量分数预测算法  

Hidden Markov model based predictive algorithm of silicon content in molten iron of blast furnace

作　　者：曾九孙[1] 刘祥官[1] 郜传厚[1] 罗世华[2] 

机构地区：[1]浙江大学系统优化技术研究所,浙江杭州310027 [2]江西财经大学信息管理学院,江西南昌331305

出　　处：《浙江大学学报（工学版）》2008年第5期742-746,共5页Journal of Zhejiang University：Engineering Science

基　　金：国家科技部重大推广项目(2005EC000166);浙江省自然科学基金资助项目(Y107110);高等学校博士点基金项目(新教师项目)(20070335161);江西省教育厅科技项目(GJJ08358)

摘　　要：为正确预测高炉铁水中硅的质量分数([Si]),提出了一种基于隐Markov模型(HMM)的预测算法.从高炉冶金反应动力学出发,分析了高炉内反应的链接关系,这种链接关系和HMM的原理是一致的.在对系统参数初始化之后,利用重估公式对参数进行训练直至收敛,从而得到系统模型.通过Viterbi算法找出所有训练样本的最大可能状态路径,并计算其似然值.将新样本输入模型得到新的状态路径及其似然值,从训练样本中找出具有相同状态路径或最小偏差似然值的序列,以训练样本下一[Si]值作为新样本下一时刻的预测值.利用该算法对高炉实际生产数据进行仿真,结果表明,与传统的人工神经网络方法相比,该方法能够有效提高预测精度和效率.A hidden Markov model （HMM） based predictive algorithm was proposed to get accurate prediction of silicon content （[Si]） in molten iron of blast furnace. According to the metallurgical reaction dynamics, chemical reactions in blast furnace have chain effect, which is similar to the principle of HMM. After the parameter initialization step, the re-evaluating formula was deployed to train the parameters using the training samples and the system model was obtained. The most likely state routes of all training samples were then computed via Viterbi algorithm and meanwhile the maximum likelihood values were obtained. New samples were input into the model and new state routes and maximum likelihood values were computed, which were then compared with those of the training samples. The most likely state route was selected, whose next [Si] was considered to be the expected prediction. The algorithm was tested using practical data. Simulation shows that compared to the traditional artificial neural networks, the HMM based algorithm can effectively improve the predictive efficiency and accuracy.

关 键 词：高炉炼铁 冶金反应动力学 隐MARKOV模型 数据模式 

分 类 号：TG250.2[金属学及工艺—铸造]