Wind power prediction is very important to the safe operation and economic dispatch of the power system

but the power prediction modeling of newly built

expanded or renovated wind farms faces two major problems: 1) New stations and station expansions cause some stations to operate insufficient data

insufficient model training. 2) Traditional shallow neural networks are difficult to cope with increasingly complex prediction input information. Therefore

a short-term wind power prediction model based on feature selection multi-level deep transfer learning is proposed in this paper. First

correlation analysis method to classify the historical data sets of other wind farms in the same province according to the correlation with the target wind farm is proposed. Then

the source wind farm prediction model to the target wind farm in the order of the correlation from low to high is migrated. Finally

the feature selection method is adopted to optimize the migration model and to ensure the effective migration of relevant features. The analysis of calculation examples shows that: 1) The multi-level deep transfer learning model can make up for the problem of insufficient training samples for new wind farms. Compared with direct modeling

the accuracy is improved by 6.5%. 2) The model optimized by the feature selection method can increase the predicted accuracy by 0.4%

so the method proposed in this paper is an effective method for power prediction of wind farms with data shortage.