Accurate short-term forecasting of meteorological variables remains a major challenge in tropical regions characterized by strong diurnal variability, seasonal transitions, intermittent atmospheric fluctuations, and highly non-stationary temporal dynamics. This study proposes a hybrid deep-learning framework combining the Stationary Wavelet Transform (SWT), Multi-Head Attention (MHA) and Long Short-Term Memory (LSTM) networks’ forecasting layers for multivariate meteorological prediction over selected locations in Togo. The proposed architecture simultaneously forecasts temperature at 2 m, relative humidity, and wind speed using hourly ERA5 atmospheric variables enriched with lagged, rolling-statistical, and cyclic temporal features. SWT decomposition extracts multi-scale atmospheric structures while preserving temporal alignment, the attention mechanism dynamically emphasizes informative temporal sub-sequences, and the LSTM layers improve medium-range temporal dependency modeling. The forecasting task was reformulated under a strictly causal prediction protocol for 24 h and 72 h forecasting horizons. The proposed framework achieved competitive forecasting performance with MAE = 2.99, RMSE = 5.22, and ${\mathrm{R}}^2=0.77$ for the 24 h horizon, and MAE = 3.44, RMSE = 6.11, and ${\mathrm{R}}^2=0.73$ for the 72 h horizon. Temperature forecasting exhibited the highest stability, whereas wind-speed prediction remained more challenging because of stronger atmospheric intermittency. The results demonstrate the relevance of multi-scale temporal representation learning for ERA5-based tropical meteorological forecasting while highlighting important methodological and operational limitations.

Accurate short-term forecasting of key meteorological variables—air temperature, relative humidity, and wind speed—remains challenging in tropical and sub-Saharan regions due to strong diurnal cycles, seasonal variability, and non-stationary dynamics. To address these limitations, this study proposes a hybrid deep learning model combining Stationary Wavelet Transform (SWT), Multi-Head Attention (MHA), and LSTM networks. First, SWT decomposes meteorological time series into multi-scale components, capturing both low-frequency trends and high-frequency fluctuations while preserving temporal resolution. Then, the attention mechanism dynamically weights the importance of these multi-scale features across time, enhancing the model’s ability to focus on the most relevant patterns and interactions. Finally, LSTM layers model long-term dependencies and nonlinear temporal structures to generate multi-output predictions. The model is trained on hourly data enriched with lagged and statistical features. Experimental results show strong predictive performance (MAE = 1.21, RMSE = 2.01, R² = 94%), with notable improvements in modeling rapid variations, especially for wind speed and humidity. This work represents one of the first integrations of SWT, attention mechanisms, and LSTM for multi-variable forecasting in tropical climates, with practical applications in energy forecasting and precision agriculture.