An Exploratory Application of Empirical Mode Decomposition and Recurrent Neural Networks for Meteorological Time Series Prediction

Abstract

Accurate short-term weather forecasting remains a critical yet complex task, particularly in tropical regions where high variability and abrupt climatic shifts can have immediate impacts on agriculture, infrastructure, and public safety. Traditional statistical methods often struggle to capture the non-linear and multi-scale nature of meteorological time series, limiting their effectiveness in localized forecasting scenarios. To address this challenge, this paper presents an exploratory prototype that combines Empirical Mode Decomposition with Recurrent Neural Networks, specifically Long Short-Term Memory (LSTM) architectures. Daily data on temperature, humidity, and atmospheric pressure from Mérida, Yucatán (2000–2018) were decomposed into Intrinsic Mode Functions, which served as input features for training separate LSTM models. The hybrid system achieved promising results, particularly for temperature and humidity, capturing key short-term patterns while highlighting limitations in pressure forecasting. These findings suggest that EMD-based preprocessing can enhance neural sequence models in dynamic forecasting contexts, offering a pathway toward more adaptive, data-driven approaches in weather-sensitive applications.