Application of a neural-network-based RLS algorithm in the generalized predictive control of a nonlinear air-handling plant

This paper presents a novel method which uses a neural network to improve the performance of a recursive least squares (RLS) parameter estimation algorithm in estimating the parameters of nonlinear processes. It also describes the use of this method in a process gain adaptive generalized predictive control (GPC) algorithm. Neural networks are well known for their ability to memorize nonlinear functions after being trained. They are often trained to represent process dynamics (or the inverse of the process dynamics) directly. This method, however, uses neural networks to learn the parameter updating process of a standard RLS algorithm and then to relate the parameters so obtained to the operating conditions of the nonlinear processes. This permits RLS to be extended to nonlinear applications. It combines the advantages of the simplicity and speed of convergence of RLS algorithms with the ability of neural networks to learn any complex nonlinear function to any desired accuracy. Experimental results when using this method to control an air-handling plant are reported and show the great potential of this algorithm in controlling nonlinear processes in which the parameters are linear but are functions of static inputs