Application of Affine Gray-Box Neural Models for Nonlinear Control of Chemical Processes

In this paper, an affine neural model is used to model the unknown part of SISO processes with un-modeled actuator dynamics. It is assumed that a partially known first-principles based model of the process, which is invertible with respect to the unknown part, is available. Using this available knowledge, I/O training data of the process, and affine neural networks, a serial gray-box model is generated which is suitable for applying feedback linearization. Hence, the resulting nonlinear controller works in a large operating region. The superiority of the gray-box over the blackbox approach is investigated for a fermentor using the experimental data borrowed from the literature. Simulation results of our case study show that the proposed affine gray-box method is superior to the conventional affine black-box method and preserves extrapolation property.