关键词: 过程先验知识, 递归神经网络, 增广拉格朗日乘子法, 粒子群-内点法优化算法, 丙烯聚合反应过程

Abstract: Most of chemical processes with nonlinear characteristics are difficult to model by general linear modeling approaches in practice.Hence, a novel recurrent neural network integrated with priori-knowledge for modeling nonlinear dynamic processes was presented.In the form of non-linear constraints, the priori-knowledge exploited from industrial chemical processes was embedded into the feed-forward neural network with the NARMAX(nonlinear autoregressive moving average with exogenous input)structure.Meanwhile, based on the augmented Lagrange multiplier(ALM)method, a hybrid PSO-IPOPT algorithm was introduced for network weight optimization.The PKRNN model with process-priori-knowledge constraint either ensured good dynamic modeling and prediction(especially extrapolation)ability, or guaranteed safety in the implementation of industrial chemical processes.The effectiveness of the PKRNN model was validated by an actual double-loop liquid propylene polymerization reaction process.

Key words: process-priori-knowledge, recurrent neural network, augmented Lagrange multiplier, particle swarm optimization-interior-point optimization, propylene polymerization reaction process