Dynamic Process Modelling using a PCA‐based Output Integrated Recurrent Neural Network

Abstract

1 C hemical processes are highly non-linear systems exhibiting complex time-dependent behavior. Modeling of these dynamics using first principle models is not always possible. As a result, alternative techniques using either data-driven or knowledge-based methods have been explored, such as statistical regression and fuzzy logic approaches. Conventional statistical regressions, however, have proven inadequate in many applications for modeling an underlying mechanism with significant non-linear characteristics, while fuzzy logic models depend heavily on expert knowledge which can be subjective. Consequently, hybrid models have recently received much attention. An interesting hybrid model proposed by Qi et al. (1999) combines first principle equations with artificial neural networks (ANN), the later being used to determine the parameters of the first principle models. The integration of fuzzy logic with principle model-based simulation was also studied by authors (Qian et al., 1999(Qian et al., , 2001) who used fuzzy logic to represent the imprecise relationships among process variables. In another effort, Baffi et al. (1999) studied the integration of polynomial spline functions with ANN.Given the large volume of data collected from computers and laboratories in modern process plants, data-driven techniques, and particularly neural networking, have become an attractive data-driven method. For modelling dynamic systems, recurrent neural networks (RNNs) need to be used (Himmelblau, 2000). While they are not suitable for high-order systems, RNNs including the Hopfield, Elman and Jordan neural networks, are effective in dealing with first order systems. To address this discrepancy, Zhu (1998) and Wu (2000) modified the Elman neural network and proposed two new approaches: a RNN with multiple subfeedback-layers, and a state-integrated RNN, respectively. The modified RNNs reserve, as much as possible, past time series information so as to facilitate the simulation of high-order dynamic non-linear behaviors.A major challenge to the use of these RNN neural networks is that for complex multiple input multiple output (MIMO) systems, the number of nodes in the input layer as well as in the hidden layer tends to be very large, resulting in training difficulties. In this contribution, a new RNN structure is proposed to address this issue, in that it integrates the past values for certain steps of the output variables with the input variables, and the original input variables are pre-processed using principal component analysis for the purpose of dimension reduction. A new methodology for modelling of dynamic process systems, the output integrated recurrent neural network (OIRNN), is presented in this paper. OIRNN can be regarded as a modified Jordan recurrent neural network, in which the past values for certain steps of the output variables are integrated with the input variables, and the original input variables are pre-processed using principal component analysis (PCA) for the purpose of dimension reduction. The main advanta...

Integrated Modelling of Process Operation Systems Using the Agent-Oriented Approach

“…It is successfully incorporated into conventional process simulators (Qian, 1999). Several efforts have been made to combine statistical analysis with non-linear regression techniques such as polynomial, spline function and ANN (Baffi et al 1999;Qian et al 2002).…”