基于极限学习的过程神经网络研究及化工应用

针对过程神经网络在化工过程建模中学习速度慢、易陷入局部极值等问题,借鉴极限学习机算法训练网络参数的思想,提出了一种新型的基于极限学习的过程神经网络(EL-PNN).ELPNN网络以过程神经网络的方式得到隐含层的输出后,不再使用梯度下降法进行参数调整,而是根据极限学习机算法通过广义逆直接求解输出权值.同时,为了进一步提高网络的泛化性能,考虑结构风险,在EL-PNN网络中加入风险比例参数.以高密度聚乙烯装置进行验证,结果表明,该网络具有学习速度快、建模精度高的特点,为过程神经网络在复杂化工生产中的应用提供了新思路.

Research and Chemical Application of Extreme Learning Based Process Neural Network

In chemical process modeling, the process neural network (PNN) usually consumes much time and falls into the local minima easily. In order to solve these problems, the extreme learning (EL) algorithm was used to train the PNN. Thus, an extreme learning-process neural network (EL-PNN) model was proposed. The outputs of the hidden layer of EL-PNN were obtained by the same means of PNN, and the weights connecting the hidden layer and output layer were then directly obtained by Moore-Penrose generalized inverse according to the EL algorithm. Meanwhile, to enhance the generalization performance of the EL-PNN, the structure risk was considered and a risk ratio parameter was introduced into the network. As a case study, the high-density polyethylene plant was selected to verify the effectiveness of the proposed model. The results show that the EL-PNN has a high learning speed and modeling precision, providing a new idea for process neural network in modeling complex chemical processes.