基于联级小波神经网络的LF精炼渣成分预报  

Prediction of LF Refining Slag Composition based on Combined Wavelet Neural Network

作　　者：张慧书[1,2] 陈韧 战东平[3] 孙丽娜 黄妍[1,2] 张作良[1,2] ZHANG Huishu;CHEN Ren;ZHAN Dongping;SUN Lina;HUANG Yan;ZHANG Zuoliang(School of Metallurgy Engineering, Liaoning Institute of Science and Technology, Benxi Liaoning 117004, China;Liaoning Key Laboratory of Optimization and Utilization of Non- associated Low- grade Iron Ore in Benxi, Liaoning Institute of Science and Technology, Benxi Liaoning 17004, China;School of Metallurgy, Northeastern University, Shenyang Lioaning 110004, China)

机构地区：[1]辽宁科技学院冶金工程学院,辽宁本溪117004 [2]辽宁省本溪低品位非伴生铁矿优化应用重点实验室,辽宁本溪117004 [3]东北大学冶金学院,辽宁沈阳110004

出　　处：《上海金属》2019年第4期80-83,共4页Shanghai Metals

基　　金：辽宁省博士科研启动基金(No.20170520079);国家自然科学基金(No.51574063、No.51874081)

摘　　要：LF炉精炼渣的成分是影响LF精炼是否达到目标的重要因素,而转炉渣成分获得是确定LF精炼渣成分的关键因素。基于神经网络有利于解决非线性问题的特点,构建了适合解决上述问题的联级预报模型。采用VB6.0进行编程,应用克服BP神经网络缺陷的小波神经网络,建立了联级小波神经网络。经研究分析确定,第1级网络结构为8×10×5,第2级网络结构为13×12×6,其中联级中的隐含层传递函数都为Morlet型函数,输出层传递函数都为S型函数。采用800炉数据进行模型训练,30炉数据现场验证表明,预报结果中32.2%炉次的绝对值相对误差在5%以内,86.1%炉次的绝对值相对误差在20%以内,最小绝对值相对误差为0,最大绝对值相对误差为33.5%。该模型预测精度较高,可以满足实际生产中对精炼渣成分预报精度的要求。LF refining slag composition is an important factor influencing whether LF refining achieves the target or not, and the acquisition of converter slag composition is a key factor in determining the composition of LF refining slag. Based on the characteristics of neural networks which is helpful to solving nonlinear problems, a combined prediction model suitable for solving the above problems was established. By using VB 6.0 software to program and application of wavelet neural network to overcome the shortcomings of BP neural network, a combined wavelet neural network was established. After research and analysis, it was determined that the first-level network structure was 8×10×5, and the second-level network structure was 13×12×6. The implicit layer transfer function was a Morlet function, and the output layer transfer function was an S function. The training data and the prediction data for the model were 800 and 30 heats respectively. The results showed that the absolute relative error of 32.2% heats was within 5%, the absolute relative error of 86.1% heats was less than 20%. The absolute relative error of minimum absolute value was 0, and the absolute relative error of maximum absolute value was 33.5%. The model had high forecast precision, it could meet the requirements for prediction accuracy of refining slag composition in actual production.

关 键 词：小波神经网络 LF精炼渣 联级预报模型 渣成分 

分 类 号：TF769[冶金工程—钢铁冶金]