Accurately predicting wind power is of key importance for large scale wind power connecting to the grid. To predict the wind speed more accurately, a combined model based on variational mode decomposition- sample entropy(VMD-SE) and bayesian neural network optimized by improved particle swarm optimization (IPSO) is proposed for ultra short-term wind power prediction. First, the wind power time series was decomposed into a series of wind speed sub-modes with different bandwidths to reduce its non-linearity by using VMD-SE. Then, the Bayesian neural network is established for all sub-modes, and the weights and thresholds of the Bayesian neural network are optimized by IPSO to obtain optimal prediction results. Simulation results demonstrate that the forecasting model based on VMD-SE has higher prediction accuracy than other conventional decomposition methods. The proposed model has higher prediction accuracy.