Avaliação da Capacidade de Generalização de Arquiteturas de Aprendizado Profundo para Previsão de Séries Temporais Meteorológicas

Abstract

The objective of climate forecasting, through time series, is to provide the most accurate results possible to assist in determining future atmospheric states. In the context of time series forecasting, the state-of-the-art focuses on deep learning algorithms and the different architectures that have emerged in recent years. However, even though complex architectures have demonstrated interesting results, the amount of work that evaluates the generalization capacity of these architectures is still tiny. Based on this, in this work, a comparative study is proposed between three neural network architectures adapted for the prediction of time series: Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN), Gated Recurrent Unit (GRU) with Attention mechanism, and U-Net convolutional architecture), aiming to predict the maximum air temperature with a one-step-ahead forecast horizon using data from Meteorology National Institute (INMET) collected across Brazil. The results showed that, although new neural network architectures are emerging, the Ensemble Learning methodology still presents better results when evaluated in the prediction of new application data, surpassing other architectures by 25% in the coefficient of determination (R²) and reducing it by up to 2.45°C of Root Mean-Square Error (RMSE).