Wind power prediction is the important basis and necessary means to solve the influence of wind power uncertainty. The requirement to prediction accuracy at each time point will be stricter with high-proportion wind power integration. Training samples are one of the key factors influencing the prediction accuracy. However

because of the complex diversity and ambiguity of the actual weather system

it is very important for improving the prediction accuracy to select the training samples that are similar to the wind conditions in the scheduling period. Therefore

a method based on wind speed cloud model is proposed to directionally select training samples with similar wind speed features in a given day. This method can improve the learning ability in capturing the randomness and fuzziness of wind speed in designated time period. By screening the historical data in an oriented way

the proposed method can enhance the prediction accuracy. Firstly

a historical database for daily wind speed cloud models is created. Then

the similarity of the wind speed cloud models is quantified

and this quantification index is used to select a sequence of historical data being similar to the wind in a predicted period. The selected samples will be used as training samples for short-term wind power prediction. In the practice

the training samples and prediction models are updated according to the weather prediction information so as to improve the prediction accuracy. The results of an example of a wind farm in Northern China are demonstrated

and verify that the proposed method can improve the accuracy of short-term wind power prediction and has practical value in engineering applications.