IEEE Transactions on Neural Networks. — Volume 1, No. 1, March 1990. — Pages 4-27.
In the literature a large variety of neural nets has been proposed all having the capability of
modeling the dynamic behavior of a system. In this paper a neural net is used to build a predictor
for such a dynamical system. This neural net is then used in a model based predictive control
algorithm.
Three types of frequently used neural networks are compared: the parallel form of the feedforward
neural network, the radial based neural network and the Elman neural network. The neural networks
are trained by using a staircase training signal instead of a random training signal.
The model validation is done by 'what f ' simulations or time-validation. The results of the model
validation test based on correlation techniques are related to the choice of the training data. Models
built with a staircase input signal are often rejected by the correlation technique because they do not
model the 'high frequency' part correctly. Simulation tests showed that the feedforward neural net
estimates the underlying non-linearity of the system frequently better than the other two networks.
Each of the neural models has been used in a predictive control algorithm for controlling the nonlinear system. This algorithm requires the minimization of a cost during the control action. The
Levenberg-Marquardt method has been used for minimizing this cost function. The robustness of the
control method is tested by adding different kinds of measurement noise and model inaccurcies.