Multi-Step Forecasting of Meteorological Time Series Using CNN-LSTM with Decomposition Methods

Disordered human actions, such as deforestation, improper land management, and industrial activities, escalate greenhouse gas emissions, contributing to global warming and climate change. Such actions amplify extreme weather events, like heat waves and storms. Hence, forecasting weather information in advance is crucial to make decisions that reduce losses and safeguard lives and property. This study evaluates the one-step and multi-step daily forecast of two hybrid approaches for various meteorological variables in three regions of Rio de Janeiro state, Brazil. The approaches include seasonal decomposition (SD), ensemble empirical mode decomposition (EEMD), Convolutional Neural Networks (CNN), and Long Short-Term Memory Neural Networks (LSTM). The effectiveness of the SD-CNN-LSTM and EEMD-CNN-LSTM models is compared to techniques such as decision tree (DT), random forest (RF), and CNN-LSTM, using both undecomposed and decomposed data (SD-DT, SD-RF, EEMD-DT, and EEMD-RF), all with parameters tuned by the Random Search method. Evaluation is performed using correlation coefficient (r), coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE). Results show that models without decomposition (CNN-LSTM, RF, and DT) are inferior. The most accurate indices are obtained by models combining real and decomposed data. The SD-CNN-LSTM model proved more suitable for one-step forecasts, while the EEMD-CNN-LSTM model is more stable for long-period forecasts. Both emerged as valuable tools for meteorological time series.