高炉铁水质量鲁棒正则化随机权神经网络建模

Robust Regularized RVFLNs Modeling of Molten Iron Quality in Blast Furnace Ironmaking

高炉炼铁过程运行优化与控制依赖于可靠、稳定的难测铁水质量(Molten iron quality, MIQ)指标模型.针对现有MIQ建模方法的不足,本文提出一种新型的数据驱动鲁棒正则化随机权神经网络(Random vector functional-link networks,RVFLNs)算法,用于实现MIQ指标在线估计的鲁棒建模.首先,为了提高建模效率和降低计算复杂度,采用数据驱动典型相关性分析方法从众多变量中提取与MIQ相关性最强的变量作为建模输入变量;其次,由于传统RVFLNs网络的输出权值由最小二乘估计获得,易受离群数据影响而鲁棒性差,引入基于Gaussian分布加权的M估计技术,提出新型鲁棒RVFLNs算法建立多元MIQ指标的鲁棒模型;同时,在鲁棒加权后的最小二乘损失函数基础上,进一步引入L1和L2两个正则化项以构成优化目标函数的Elastic net,用于稀疏化RVFLNs网络的输出权值矩阵,解决RVFLNs网络多重共线性和过拟合的问题.最后,基于某大型高炉工业数据,进行充分数据实验,结果表明所提方法具有更高的建模与估计精度以及较强的鲁棒性能.

Optimal operation and control of a practical blast furnace(BF) ironmaking process depend largely on a reliable model of molten iron quality(MIQ) indices that can not be measured online. Aiming at the shortcomings of the existing MIQ modeling methods, a new data-driven robust regularized random vector functional-link networks(RVFLNs) algorithm is proposed to realize robust modeling of MIQ indices. First, to improve modeling efficiency and reduce computational complexity, the data-driven canonical correlation analysis(CCA) is employed to identify the most influential components from multitudinous factors that affect the MIQ indices to serve as the input variables. Next, since the output weights of traditional RVFLNs are obtained by the least squares approach, the robustness may decrease when the training dataset is contaminated with outliers. To solve this problem, the robust RVFLNs model of MIQ using Gaussian distribution weighted M-estimation is established. Simultaneously, on the basis of the least-square loss function of the robustness,the L1 regularization and L2 regularization are introduced to achieve sparse output weight and prevent the overfitting and multicollinearity of the RVFLNs model by forming the Elastic net that optimizes the objective function. Finally,experiments using industrial data from a large balst furnace have demonstrated that the proposed method produces a higher modeling, estimating accuracy and stronger robustness than other modeling methods.