Neural based predictive control of a multivariable microalgae fermentation

This paper discusses the application of a multivariable nonlinear model predictive control (MPC) scheme where the model is given by a neural type nonlinear autoregressive with exogenous input (NARX) model. The simulated microalgae fermentation process presented here shows a nonlinear and time-varying behaviour and provides an interesting example of modelling and controlling complex systems. We show how to identify the neural NARX model, how to incorporate the developed model into the formulation of predictive control, and how neural based predictive control (NPC) scheme can work well. The optimisation algorithms used in neural training and predictive control are the conjugate gradient algorithm with exact line search and a modified Maquardt´s algorithm with coarse line search respectively. A series of simulation studies, including the comparison with differential geometry techniques and predictive control based on adaptive polynomial NARX model, demonstrate a satisfactory control behaviour