Hybridized GA-optimization of neural dynamic model for nonlinear process

Abstract

Neural networks as universal approximators possess capability to model complex nonlinear phenomena. However, when almost nothing is known about the modeled dynamic process it is difficult to determine important parameters like the number of neurons or the size of regressor vector (dynamic order). In order to avoid suboptimal settings for a dynamic model using trial-and-error method, genetic algorithm is used for optimizing the neural dynamic model. To improve the results even more, the genetic optimization is hybridized with a local optimizer in the form of Levenberg-Marquardt algorithm commonly used for neural network training. Here a neural model of biomass-fired boiler emissions is considered, which is eventually intended for predictive control. Series-parallel NARX model is used with two hidden layer neural network and tan-sigmoid transfer functions. The simpler neural model structure will be computationally less expensive what is important for online predictive control. The results confirm the capability of this method to achieve simpler network structures with errors comparable to the case when trial-and-error settings were previously used.