Short-term wind speed is intermittent, fluctuating, nonlinear and non-stationary, and has a high degree of complexity, which is difficult to predict. The wind speed signal can be seen as coupled from simple signals with low complexity and strong regularity, so the decomposition method can be used to make it divided into multi-scale fluctuating energy, reduce the component complexity and enhance its regularity, which can improve its prediction accuracy. Therefore, to improve the learning efficiency of the neural network, the Kmeans algorithm is used to cluster the original wind speed data on similar days. Secondly, the wind speed sequence is decomposed using VMD to extract the multiscale regularity. Finally, because the LSTM neural network is more capable of capturing the fluctuation regularity of the long-dependent sequence, the decomposed wind speed components are predicted by using the LSTM neural network, and the final prediction results are obtained by superimposing the predicted values of each component. The combination prediction model based on Kmeans-VMD-LSTM can effectively improve the accuracy of short-term wind speed prediction as shown by a large number of experiments and comparisons between different methods.