The GMR Neural Network for Inverse Problems

This paper deals with the mapping approximation problem by means of a neural network. In particular it presents the GMR (Generalized Mapping Regressor) neural network, which belongs to the family of self-supervised NNs. It is an incremental self-organizing neural network which can approximate every multidimensional function or relation presenting any kind of discontinuity. It can also simultaneously compute the inverse of any function to be approximated, if it exists. In this paper, GMR is used in inverse modeling for the control of a PEM fuel cell stack. In particular the output voltage of the PEM-FC, which is a non linear system, is controlled. A new control scheme based on the GMR has been developed, called PID-GMR, which adopts the scheme of Kawato (1990). The PEM-FC inverse model created by the GMR is added to a classic PID regulation system. The simulations show that the PID-GMR scheme outcomes the classical PID control with particular regard to the steady-state accuracy.