Model predictive control using neural networks

In this article, we present the application of a neural-network-based model predictive control scheme to control pH in a laboratory-scale neutralization reactor. We use a feedforward neural network as the nonlinear prediction model in an extended DMC-algorithm to control the pH-value. The training data set for the neural network was obtained from measurements of the inputs and outputs of the real plant operating with a PI-controller. Thus, no a priori information about the dynamics of the plant and no special operating conditions of the plant were needed to design the controller. The training algorithm used is a combination of an adaptive backpropagation algorithm that tunes the connection weights with a genetic algorithm to modify the slopes of the activation function of each neuron. This combination turned out to be very robust against getting caught in local minima and it is very insensitive to the initial settings of the weights of the network. Experimental results show that the resulting control algorithm performs much better than the conventional PI-controller which was used for the generation of the training data set