To improve the prediction accuracy of wind power output, a model using a long- and short-term memory (LSTM) artificial neural network based on Bayesian optimization is proposed. First, empirical mode decomposition is used to decompose an historical output series of wind power, and eight statistical features are extracted from each component and the original data respectively. In this way a prediction feature set is formed together with the output values at the first six moments of prediction. Then, a least absolute shrinkage and selection operator (LASSO) algorithm is used to extract the feature subset with statistical significance from the prediction feature set as the input to the prediction model. Finally, an optimization method of an LSTM network based on Bayesian super-parameters optimization is proposed to improve prediction accuracy. The historical data of wind power output in a city in Hubei Province are selected for a prediction experiment. The results show that, compared with a BP neural network, support vector machine (SVM), radial basis function (RBF) network, general regression neural networks (GRNN) network and other prediction models, the proposed model has higher accuracy and the feature extraction method is more reasonable.