Long Short-Term Memory Approaches for Weather Forecasting from Local Stations

Weather forecasting is a relevant task affecting human activities, including agriculture, transportation, economy, environment, tourism, and entertainment. The weather conditions are predicted based on complex mathematical models that require considerable computer resources and data from satellites, sensors, and atmospheric simulations. The successful application of recurrent neural networks for weather forecasting worldwide motivated us to explore these approaches on weather data from the National Institute of Meteorology (Instituto Nacional de Meteorologia - INMET), in Brazil. The goal of this research is to compare different recurrent neural network techniques for weather time-series forecasting using data from weather stations located in Brazil in the cities of Brasilia, Florianopolis, and Manaus. Weather forecasting is handled as a regression task since the prediction of the next weather conditions depends on previous atmospheric parameters. The proposed methodology consists of several steps including data collection, preprocessing, sampling via hold-out, hyperparameter optimization, regression experiments, evaluation, and discussion. The Bidirectional Long Short-Term Memory (BiLSTM) presented the best overall results for the regression task compared to Long Short-Term Memory (LSTM). The source code, datasets and resources are available at https://gitlab.com/gvic-unb/icfsp2024lstmweather-inmet.