Combinatorial Network of Dynamic Models: A Method to Improve Bad-quality Models

This work proposes a combinatorial network of dynamic models in order to combine models that doesn't approximate well a system, to obtain an output that have a better performance considering the system behaviour. The network can be useful in situations where a good model cannot be obtained from data, such when there is a bad signal noise ratio in identification data-set or when a specific input cannot be applied to generate identification data. To do this, we present two different approaches: an analytical one and a numerical method, both combining a weighted sum of the bad quality models. The method is tested on models obtained through a multi-objective system identification procedure, and from models obtained through an interval system identification procedure. The combined model has improved the performance in the validation indexes analyzed, reaching a reduction up to 65% in the RMSE index, 95% in the MSE index of the static curve, 87% in the energy of the residues vector and a reduction of 21% in the auto-correlation energy of the residues vector.