Dynamic neural networks partial least squares (DNNPLS) identification of multivariable processes

This paper presents the dynamic neural networks partial least squares (DNNPLS) as a strategy for open-loop identification of multivariable chemical processes that circumvent some of the difficulties associated with multivariable process control. The DNNPLS is an extension of the neural networks' partial least squares (NNPLS) developed by Qin and McAvoy (Comp. Chem. Eng. 20 (1992) 379). Here, a dynamic extension to the NNPLS algorithm is proposed in which the static neural network models in the latent space (inner relationship) are replaced by dynamic neural network models. Though this approach has previously been dismissed as being sub-optimal (Am. Inst. Chem. Eng. J. 38 (1992) 1593; Chem. Eng. Sci. 48 (1993) 3447) in terms of the outer relationship (relationship between the residuals), Lakshminarayanan et al. (Am. Inst. Chem. Eng. J. 43 (1997) 2307) have shown that this sub-optimality problem comes into prominence only when no attention is placed on the design of the plant probing signals. As illustrations, the DNNPLS identification strategy is implemented on simulations of a model IV fluid catalytic cracking unit (FCCU) and of an isothermal reactor. In both cases, it is shown that the methodology is capable of modeling the dynamics of the chemical processes and an improved performance is achieved over that of the PLS-ARMA (Comp. Chem. Eng. 20 (1996) 147) for the isothermal reactor. (C) 2002 Elsevier Science Ltd. All rights reserved.