Integrating Dynamic Neural Network Models with Principal Component Analysis for Model Predictive Control ⁎ ⁎ Financial support from the Natural Sciences and Engineering Research Council of Canada (NSERC) and the McMaster Advanced Control Consortium (MACC) is gratefully acknowledged.

This work addresses the problem of identifying models using process data with possibly correlated manipulated variables for model predictive control (MPC) design. The key idea is to use principal component analysis (PCA) to reject the redundancy in the input space and utilize scores to build the dynamic model of the system using recurrent neural networks (RNN). The identified PCA-RNN model is then used in the MPC optimization problem, calculating the optimal scores. The control actions are computed using the loadings of the PCA model. The efficacy of the proposed approach is evaluated using a chemical reactor example. The results are compared with a base-case scenario where the data is directly used to build a dynamic neural network model and used as part of a model predictive control implementation. The simulation results show the superiority of the proposed integrated PCA-RNN models for model predictive control.