In order to improve wind power prediction accuracy, an ultra-short-term wind power prediction model based on the combination of two-layer decomposition technique and particle swarm optimization long short-term memory (PSO-LSTM) neural network is proposed. The fast ensemble empirical mode decomposition (FEEMD) method is used to deconstruct the original wind power sequence into a series of intrinsic mode function (IMF) components and the remainder term. The high frequency IMF is decomposed by the variational mode decomposition (VMD) of two-layer decomposition technology. The sample entropy is used to solve the problem of too many components and complicated calculation. The input dimension is determined by selecting the elements of high correlation degree with the predicted value through partial autocorrelation coefficient function (PACF). PSOLSTM is used to construct the prediction model, and the final value is obtained by superposition. The experimental results show that the combined model can effectively improve the prediction accuracy.