带有稳定学习的递归神经网络动态偏最小二乘建模

针对目前非线性动态偏最小二乘(PLS)建模方法在拟合较强非线性化工过程时存在的问题, 提出一种基于稳定学习的递归神经网络动态PLS建模方法. 该算法将递归神经网络与Hammerstein模型相结合, 对外部PLS提取的特征向量进行内部建模, 具有逼近较强非线性化工过程的能力, 改善了模型的适用范围. 此外, 采用带有稳定学习的参数更新算法对模型参数进行在线修正, 改善了模型的预测精度和自适应能力. 将此方法应用于氧化铝生产过程铝酸钠溶液组分浓度建模实验, 仿真结果表明, 本方法是可行有效的.

Dynamic partial least squares modeling with recurrent neural networks of stable learning

A dynamic modeling algorithm is proposed for a strongly nonlinear chemical process, it is based on partial least squares (PLS) and recurrent neural networks with a stable learning rate. The outer PLS algorithm reduces the dimensionality of data and extracts score vector, and the inner model which combines recurrent neural networks with Hammerstein model captures the nonlinear characters to extend the model application scope. Besides, the stable learning algorithm updates the model parameters to improve the prediction precision and adaptation ability. This method is implemented in the process of alumina production to measure the component concentrations of sodium aluminate solution. Simulation results show that the modeling method is effective.