Modelagem de temperatura máxima não estacionária de Piracicaba (SP): uma abordagem bayesiana

Abstract

Temperature is one of the phenomena that have been studied and debated by many researchers about the increases (trends) perceived in the intensities of temperatures in the last decades, and the occurrence of maximum temperatures, which is known as an extreme event originated mainly by global warming, and that they are related to the low probability of occurrence (rare events), and being of high impact due to the immense losses that they can cause. The theory of extreme values plays a fundamental role in modeling rare events that have become a topic of notable relevance in recent decades. The present study aims to model, via Bayesian inference, the maximum temperature of Piracicaba-SP using linear and non-linear functions to incorporate the trend in the position parameter of the generalized distribution of extreme values (GEV). Normal and uniform distributions were used as prioris to elicit information. The obtaining of the marginals of the distributions a posteriori was used the Monte Carlo method via Markov chains. The best distributions were selected by analyzing the Bayes factor and decision criteria, Deviation Information Criterion (DIC) and the mean prediction error (EMP). The GEV distribution incorporating a non-linear trend in the position parameter proved to be efficient in forecasting the maximum monthly temperature (TMM) of Piracicaba-SP and presenting more accurate, reasonable results and with smaller forecasting errors, it was found that in the months January to March and August and September, the best distribution was stationary GEV. In the months of April to June and October, the best distribution was non-stationary GEV with a linear trend and with a non-linear trend for the months of July, November and December. In four of the informative a priori months it provided more accurate results. The maximum temperatures expected for different return times were predicted.