Direct Adaptive Control of Process Systems Using Recurrent Neural Networks

One of the main draw-backs of the current adaptive, as well as model predictive, control schemes is that they are designed using linear or linearized system models. A method for the adaptive control of non-linear and non-minimum phase plants using recurrent neural networks is proposed, based on model predictive control concepts. A conventional PI (proportional+integral) controller structure is used for the initial simulations. A recurrent multilayer perceptron network is used for offline and on-line system identification of the plant, while a steepest descent learning algorithm is used to estimate the empirical model parameters such that some modeling related objective function is minimized. Similarly using steepest descent, the gains of the controller are varied so as to minimize an alternate control related error criterion, such as the tracking or regulation error in a finite horizon. A U-tube steam generator (UTSG) is an ideal example of a non-linear, non-minimmum phase system. A piece-wise linearized model of the UTSG, which captures the dynamics of the actual model to sufficient accuracy, is used for testing the proposed control algorithm.