Abstract: Driven by the national strategic goal of “carbon peak, carbon neutrality”, wind power generation is an important part of the clean energy system, and the improvement of its power prediction accuracy has become a key issue to be solved in the field of new energy. At present, the main constraints affecting the prediction accuracy of wind power include two aspects: one is the outlier problem in the process of data collection, and the other is the influence of meteorological conditions on the power generation. To solve this problem, a comprehensive solution is proposed. In the data preprocessing stage, the outlier detection method combining quartile method and isolation forest(iForest) algorithm is adopted. In the aspect of feature selection, ReliefF algorithm and minimal redundancy maximal relevance(mRMR) method are combined to realize scientific screening of multi-dimensional meteorological features and effectively reduce the data dimension. Finally, the accuracy of ultra-short-term wind power prediction is significantly improved by constructing a prediction framework based on the combined model. Meanwhile, the data cleaning, decomposition and dimensionality reduction methods proposed in this paper are validated. The simulation results show that the method proposed in this paper has a good effect on improving the prediction accuracy and has certain feasibility.