Aplicação do método enxame de partículas na otimização de modelos de regressão multivariada aplicados na quantificação da mistura biodiesel/óleo vegetal/diesel.

Abstract

The gradual substitution of products coming from non-renewable sources by products of renewable origin has been globally discussed lately, especially regarding to the fuels. In Brazil, the government has declared official the biodiesel/diesel mixture in the obligatory percentile of 2% starting from 2008. This dissertation based on the need for monitoring such mixture’s quality proposes the study and the use of the particles swarm method (Particle Swarm Optimization - PSO), together with the interval partial least-squares multivariate regression technique (iPLS). It presents as its aims to constitute a hybrid algorithm capable to select more representative variables of the representative to be quantified, seeking to produce more robust regression models. In this research, it was done the biodiesel, the vegetable oil, and the diesel oil quantification, using (a) the whole spectral area of the mixtures, (b) the spectral areas suggested by the norm proposal from the Brazilian Association of technical rules (ABNT), and (c) the areas chosen by the hybrid algorithms. In the biodiesel quantification, the smallest foreseen mistake obtained was through the hybrid method, being this mistake 61.54% smaller than the mistake found by the areas suggested by the norm proposal from ABNT and, 26.72% smaller than the found mistake using the whole spectral area of the mixtures. Yet in the vegetable oil quantification the smallest foreseen mistake, also obtained through the hybrid method, was 27.64% smaller than the mistake found by the areas suggested by the norm proposal from ABNT and, 38.62% smaller than mistake found using the whole spectral area of the mixtures. However, in the diesel oil quantification the methods have presented similar performance. Based on these results, it was possible to conclude that the use of particles swarm method in association with the method iPLS, was capable to optimize the solutions, selecting the spectrum’s variables for the construction of more robust models.