聚合物分子量分布(MWD)是反映产品性能最重要的指标之一,它是典型的二元建模对象,采用组合神经网络对MWD的空间和时间变量进行分解建模.首先利用离散正交多项式神经网络在链长空间上建立分布与链长的模型,然后将MWD与时间变量的关系转换为网络权向量与输入变量之间的函数,利用递归神经网络建立两者之间的模型,最后组合两个网络达到建模目标.分布函数的模型表达式可写成状态方程形式,为进一步设计控制策略提供了基础.在链长空间上建立模型时,实现了神经网络的权向量与MWD相应阶次矩值之间的等价关系,网络权向量由单纯的拟合数据转变为有意义的物理量,实现了神经网络模型的灰箱化,为精确预测网络隐层节点数问题提供了解决途径.提出的方法应用于实验室规模的苯乙烯聚合过程,证明了建模方法的可行性,同时网络权值与矩值的等价关系也得到验证.%The molecular weight distribution ( MWD) of polymer is one of the most important performance indexes, which is a kind of typical binary modeling. A method based on hybrid discrete orthogonal polynomial neural network (DOPNN) was proposed to model the MWD of polymers. First, the space and time variables were decomposed by the hybrid neural network. The DOPNN was used to obtain the space model of MWD, and the relationship between MWD and input variables (namely the time variables) was converted into the function between weight vector of space model and input variables. Second, the recurrent neural network was used to obtain the time model. Last, the modeling destination was reached by combining the two NN models mentioned above. The mathematical expression of the model was similar with the traditional discrete state-space expression. Based on the model, an easy way to design the control strategy could be achieved. In space modeling, the weight vector of NN was equivalent to the moment of MWD. That is to say, the weight vector of neural network was of practical significance, so that the grey box model could be obtained. A solution to forecast the number of hidden nodes of neural network was provided. The experimental system investigated was the styrene polymerization in CSTR. The results of the experiment indicated that the NN model was able to capture the MWD as well as to provide accurate moment of MWD through the weight vector of NN model.
展开▼
