Modelling of a motor compressor group feeding a hydrogen fuel cell using recurrent neural networks

Artificial recurrent neural networks (ARNN) are well known as effective modelling tools for complex dynamic systems. However, easily implementing training algorithms such as back propagation through time and quasi-Newton algorithms require great amounts of memory and computational resources; moreover, those grow geometrically with the size of the time series (number of points in the sample used as training example). In the case of multiple output systems, the complexity of these algorithms grows also in a geometrical form, related to the number of system outputs and the number of hidden layers in the network, thus increasing the computation time to obtain a satisfactory solution. In this paper, some methodological and architectural solutions are proposed to overcome this drawback. The implementation simplicity of back propagation and quasi-Newton algorithms is nevertheless maintained, only their memory consumption and computation time is optimized to obtain a satisfactory dynamical model of a motor compressor group feeding a hydrogen fuel cell (in fact, a multiple input, multiple output system). To accomplish this goal, an artificial feedback loop is implemented and a parallel structure of one-output networks is selected as network architecture. The obtained results show that this architectural and methodological approach can be an interesting training option for complex multiple output system modelling using recurrent neural networks.