Uso de informação auxiliar em redes neurais e formação de comitês na identificação de sistemas dinâmicos

Abstract

This work is dedicated to study two strategies for identification of nonlinear dynamical systems, specially, when available data are not representative. A procedure to incorporate auxiliary information during the training of a neural network and a study about combining dynamic models are presented. The background of both approaches and experimental applications are shown. The first approach (gray-box identification) uses static curve as auxiliary information, which is added in the neural network parameters adjustment stage (training). A multi-objective approach is adopted to minimize both free-simulation error on dynamic data and static curve error. The proposed technique was applied to two experimental processes: a pilot hydraulic pumping system and an industrial gas-lift offshore oil well. The proposed gray-box identification technique was compared to black-box approaches, particularly, in operating regimes that were not available in the dynamical identification data sets. Results shows that the gray-box procedure yields models with better performance than the ones obtained by the black-box approach in, at least, one of the objectives: static function error or dynamic test data prediction error, where the dynamic data cover a broader operating range. It is shown that the implementation of this gray-box approach is justifiable when the black-box procedure does not achieve a model with good static performance. The second approach, on combining dynamic models, is presented with focus on diversity concept. Techniques to measure diversity are described, specially those without restrictions about model class or structure, i.e. which uses only input and output data sets. Simulated problems shows strong influence of data set on this kind of diversity metrics. Based on the concept, a procedure to measure diversity is proposed, which specifies appropriate data properties for the measurement. Using this measure, a weighted average model combiner is proposed and applied on an industrial gas-lift offshore oil well. Results shows that combiner can reduce the value and variance of estimation error, in comparison with simple averaging, specially when models diversity are not assured.