A comparison of nonlinear techniques for the identification of dynamic models

A variety of nonlinear dynamic modeling techniques were tested on data from a real pilot scale apparatus. The models were tested for their ability to model the system, interpolate and extrapolate to regions outside the modeling set, and to function in the presence of noise. An artificial neural network with linear terms was found to perform as good or better than the other techniques on this data. An input transformation method using a genetic algorithm to select the transformation structure and regression to determine the parameters was able to perform equally well with less computation effort. The nonlinear biased regression techniques did not perform as well though they required less computation and the results were reproducible