Due to the limited reserves of traditional energy and the excessive use of non-renewable energy， the global carbon emissions seriously accelerate the global greenhouse effect. Compared with traditional energy， wind energy is a kind of green renewable energy. Accurate wind speed prediction is of great significance to wind power generation system， but the inherent randomness of wind speed signal makes its fluctuation complex and uncontrollable. It is difficult to grasp the characteristics of wind speed series using single or fixed hybrid models in previous studies. Therefore， a wind speed forecasting system is proposed based on decomposition and the optimal machine learning models. First， the Variational Mode Decomposition （VMD） algorithm is used to decompose the original wind speed sequence into several sub-sequences. Then， five models of Fuzzy Neural Network （FNN）， Nonlinear Auto-regressive Neural Network （NARX）， Elman neural network， Back Propagation Neural Network （BPNN） and Auto-regressive integrated Moving Average model （ARIMA）， are selected to construct a machine learning model selection system and predict the validation set of sub-sequences， respectively. The root mean square error （RMSE） is used to select the optimal model and the final prediction result is obtained by summing up the predicted values of each sub-sequence. The measured wind speed data from four stations in Ningxia Hui Autonomous Region was used in the experiment. The sampling height was 70 meters and the time interval was 15 minutes. Three groups of comparison models were established to verify the prediction effect of the proposed model. The first group： five single models used in this experiment： FNN， NARX， Elman， BPNN and ARIMA； The second group： five hybrid models based on VMD decomposition： VMD-FNN， VMD-NARX， VMD-Elman， VMD-BPNN and VMD-ARIMA； The third group： newer hybrid models of deep learning： Improved Complete Ensemble Empirical Mode Decomposition （ICEEMDAN）-Long Short-Term Memory （LSTM） and ICEEMDAN-Gated Recurrent Unit （GRU）. The experimental results indicate that the proposed model is superior to other baseline models. The VMD decomposition method can effectively reduce the complexity of the original wind speed sequence. For different sub-sequences after decomposition， the optimal method of machine learning model can capture the sequence features and obtain the optimal prediction results.