Convolutional neural network is designed for probabilistic wind power forecasting. Ensemble technique is used to cancel out the diverse errors of point forecasters. The model misspecification and data noise in wind power are separately evaluated. The competitive performance and robustness of the proposed method were proved.