RBF-CSR方法及其应用于裂解装置建模的研究

RBF－CSR是在分析RBF－PLS的基础上提出的新方法。它保留了RBF－PLS的优点：采用神经网络的结构，又用数学方法直接求解，免去了ANN冗长的训练过程和其它诸多欠缺。RBF－CSR方法可以在更宽广的空间内寻找最优的网络参数，它所建立的模型具有很强的预报精度和良好的稳定性，又有简洁的解析形式，便于优化等进一步的计算和处理。该方法已成功地应用于裂解装置的建模。

The Radial Basis Functions-Cyclic Subspace Regression Approach and its Application to Cracker Modeling

Artificial neural network is a frequently-used modeling method in chemical engineering, especially for problems with complex mechanism. But the training of ANN is difficult. A lot of trouble is generated in the training process such as overfitting, converging to local optimum, etc. A new idea RBF-PLS approach that artificial neural network and regression method are combined for solving the problem was presented by Massart. In this article the RBF-CSR approach was proposed by analyzing RBF-PLS. The approach has the merit of RBF-PLS, i.e. using a structure similar to that of neural network, getting solution by mathematical methods directly, without the tedious training process of ANN and other evoking shortcomings. Finding the optimal coefficient in wider space, RBF-CSR can improve the accuracy and stability of predicted value of model. Thus models created by RBF-CSR were better than that created by RBF-PLS. Moreover the models had a brief analysis formula which was convenient for further processing such as optimization. It was successfully applied to cracker modeling.