Análise comparativa entre modelos de regressão distribucional e os principais algoritmos de aprendizado de máquina na predição de dados meteorológicos

Abstract

Univariate regression models date back to the 19th century and aim to comprehend how a set of explanatory variables influences or explains a response variable. While it is common to encounter papers comparing flexible machine learning methodologies with conventional regression models, such a comparison may not be suitable due to the stringent assumptions and limited flexibility of typical regression models. Therefore, this dissertation proposes to assess and compare the performance of distributional regression models, initially proposed as generalised additive models for location, scale, and shape (GAMLSS), which represent a more modern and flexible approach, with other commonly employed machine learning algorithms in the literature, namely: random forest, support vector regression, extreme gradient boosting, and prophet, for meteorological datasets. In our first article, already published in a journal, the need to use GAMLSS in modelling daily average temperature over a one-year period in the city of Florianópolis, Brazil, was emphasized. This study demonstrated that less complex regression models would not be suitable for fully explaining the response due to the different regression structures built into its distribution. In the second paper, we compare the predictive performance of GAMLSS with the four other mentioned machine learning algorithms. We used data from an automatic weather station in the city of Florianópolis, Brazil, collected over 10 years (from 30 March 2013 to 28 March 2023). GAMLSS based on the Box-Cox t distribution returned more satisfactory results in most metrics used for comparing the fitted models, proving to be an interesting and robust alternative for fitting and predicting meteorological data.