<p>The accurate prediction of short-term wind speed plays a crucial role in the early warning and regulation of wind farms, enabling effective power generation planning, optimizing power dispatch, and ultimately reducing carbon emissions. To improve prediction accuracy, it is essential to address the nonlinearity and non-stationarity present in wind speed time series data. In this study, we propose a hybrid machine learning framework for short-term wind speed prediction using historical wind datasets. The framework begins by employing an improved signal decomposition technique called variational mode decomposition to decompose the original wind sequence into intrinsic mode functions. This decomposition helps to capture the underlying patterns and dynamics of the wind speed data. Subsequently, we reconstruct the subsequences using sample entropy, resulting in smoother and more periodic subsequences that are better suited for prediction. Finally, a short- and long-term memory neural network is utilized for sequence prediction based on the reconstructed subsequences. Our analysis demonstrates that incorporating the sequence provided by signal decomposition technology as input significantly improves prediction accuracy compared to traditional mainstream models. Moreover, when dealing with high-dimensional sequence subsets generated from large-scale wind history datasets, the selected sequence subset not only reduces dimensionality but also improves prediction accuracy by at least 26% according to various evaluation criteria. Therefore, the method proposed in this study has the potential to accurately predict short-term wind speed using extensive wind history datasets. This has significant implications for wind resource management and wind turbine regulation and early warning systems.</p>