On the Structural Optimization of a Neural Network Model Predictive Controller

Abstract

This paper provides a new algorithm for tuning the two most effective parameters in nonlinear model predictive control (NMPC). Tuning is performed in two steps. First, a new method based on Barron’s formula, the bicoherence nonlinearity test, and inphase-quadrature demodulation is proposed to determine the number of hidden layer neurons in a two-layer neural network. In the second step, a fuzzy algorithm is introduced to tune the input weight matrix in the objective function to make the tuning problem more practical and precise. To show the effectiveness of the proposed method, several examples are discussed including a simple flow process and a more complex pH neutralization problem. The method is also evaluated in the laboratory scale pressure and level processes. It is shown that the proposed method leads to tuning the number of neurons and the weight matrix with an acceptable performance.