PREDICTING RELATIVE HUMIDITY USING UNIVARIATE MODELS: A COMPARISON BETWEEN BILS™ AND HYBRID ROLLING VMD-BILS™

Abstract

This article evaluated the performance of univariate relative humidity forecasting using a traditional BiLSTM deep learning model versus a hybrid Rolling VMD-BiLSTM model. The methodology employed was based on the use of daily meteorological data from eight municipalities in Paraná, Brazil (2007 to 2023). Cyclical encoding was applied to handle seasonality, and the Optuna framework was used for hyperparameter optimization. The main methodological differential was the application of Variational Mode Decomposition in a sliding window (Rolling VMD), processing the time series iteratively to mitigate data leakage, ensuring a realistic predictive scenario. The main results showed the superiority of the hybrid approach, which reduced the RMSE error by 14.63% and increased the explained variance ( ) by 32.35% compared to the baseline model. Furthermore, the hybrid model mitigated the time lag and the underestimation of climatic extremes observed in the pure network. It is concluded that the coupling of the Rolling VMD algorithm to the BiLSTM network constitutes a rigorous and reliable algorithmic alternative, preventing anticipation biases and offering robust support for continuous monitoring systems and the issuance of meteorological alerts.