LSTM approach for short-term prediction of precipitation values in Rio Grande do Sul

- 337600
Resumo
Favoritar este trabalho
Como citar esse trabalho?
Resumo

Extreme climatic events have become increasingly frequent, particularly in southern Brazil. In May 2024, the state of Rio Grande do Sul experienced one of the most severe flooding events in its history after weeks of intense precipitation. Motivated by this scenario, this work investigates the use of machine learning techniques for short-term rainfall prediction. We apply a Recurrent Neural Network (RNN) approach, focusing on the Long Short-Term Memory (LSTM) architecture, to model daily precipitation time series extracted from the ERA5 atmospheric reanalysis dataset. The analysis uses the ERA5 grid cell corresponding to the city of Porto Alegre and considers sequential data from 2016 to 2020. Several model configurations were tested by varying parameters such as sliding window size, loss functions, and data scaling techniques. Three representative models were selected to characterize different rainfall regimes (low, medium, and heavy precipitation). In this work, a mixed configuration combining these models is proposed.

Compartilhe suas ideias ou dúvidas com os autores!

Sabia que o maior estímulo no desenvolvimento científico e cultural é a curiosidade? Deixe seus questionamentos ou sugestões para o autor!

Faça login para interagir

Tem uma dúvida ou sugestão? Compartilhe seu feedback com os autores!

Instituições
  • 1 Universidade Federal do Rio Grande do Sul
Eixo Temático
  • ST08 - Modelagem Computacional
Palavras-chave
Neural networks
Climate change
Timeseries forecasting