Wavelet-Based Hybrid Model for Short-Term Wind Speed Forecasting in Coastal Brazil

Short-term wind speed forecasting remains challenging due to the inherent nonstationarity and multiscale variability of atmospheric processes. This study proposes a hybrid multiscale forecasting framework combining four-level discrete wavelet transform decomposition with scale-specialized predictive models and an adaptive Self-Error Stabilization Mechanism. Unlike conventional wavelet-machine learning hybrids, which do not explicitly address reconstruction-stage error propagation, the proposed framework introduces a causal, scale-aware attenuation operator that suppresses inconsistent high-frequency predictions before inverse reconstruction. Low-frequency dynamics are modeled using a Long Short-Term Memory network, while medium- and high-frequency components are predicted by dedicated machine learning regressors. Evaluated on a 22-year daily wind speed dataset from the Brazilian coast (INMET, 2001&#x2013;2023), the LSTM&#x2013;MLP configuration achieved an RMSE of 0.275 m/s, MAE of 0.188 m/s, sMAPE of 1.99%, <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> of 0.991, and C30 of 99.97% at the one-step-ahead horizon. Structural ablation analysis confirms the importance of explicit reconstruction control, as removing SESM increased RMSE from 0.275 to 1.350 m/s and reduced <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> from 0.991 to 0.784. These results indicate that reconstruction-aware multiscale modeling improves forecasting stability and physical consistency while preserving controlled structural complexity, providing a computationally efficient framework for wind farm operation and grid management. When benchmarked against comparable studies reported in the literature in comparable daily t+1 wind speed forecasting studies, the proposed framework can yield up to a 47% reduction in RMSE.