Predicting temperatures in Brazilian states capitals via Machine Learning

TL;DR

This study uses Random Forest machine learning to accurately forecast monthly temperatures in Brazilian capitals from 1961 to 2022, considering emissions, past temperatures, and their combinations, revealing regional climate change patterns.

Contribution

The paper introduces a novel approach combining emissions and historical temperature data in RF models for precise temperature forecasting in Brazil.

Findings

High forecast accuracy with NMRSE less than 0.083.

Regional differences in temperature trends and emissions identified.

Best forecast region is Northeast Brazil with NMRSE = 0.012.

Abstract

Climate change refers to substantial long-term variations in weather patterns. In this work, we employ a Machine Learning (ML) technique, the Random Forest (RF) algorithm, to forecast the monthly average temperature for Brazilian's states capitals (27 cities) and the whole country, from January 1961 until December 2022. To forecast the temperature at $k$-month, we consider as features in RF: $i)$ global emissions of carbon dioxide (CO$_2$), methane (CH$_4$), and nitrous oxide (N$_2$O) at $k$-month; $ii)$ temperatures from the previous three months, i.e., $(k-1)$, $(k-2)$ and $(k-3)$-month; $iii)$ combination of $i$ and $ii$. By investigating breakpoints in the time series, we discover that 24 cities and the gases present breakpoints in the 80's and 90's. After the breakpoints, we find an increase in the temperature and the gas emission. Thereafter, we separate the cities according to their geographical position and employ the RF algorithm to forecast the temperature from 2010-08 until 2022-12. Based on $i$, $ii$, and $iii$, we find that the three inputs result in a very precise forecast, with a normalized root mean squared error (NMRSE) less than 0.083 for the considered cases. From our simulations, the better forecasted region is Northeast through $iii$ (NMRSE = 0.012). Furthermore, we also investigate the forecasting of anomalous temperature data by removing the annual component of each time series. In this case, the best forecasting is obtained with strategy $i$, with the best region being Northeast (NRMSE = 0.090).