Abstract: Short-term wind power forecasting is crucial for real-time dispatch of power systems. Reliable wind power forecasts not only ensure the safe operation of power system but also enhance the operational efficiency of power grid.To obtain more accurate and reliable wind power forecasting results, considering the nonlinear and time-series characteristics of wind power data, a short-term wind power forecasting method based on feature similarity day combined with CNN-BiLSTM-Attention was proposed. Firstly, the impact of meteorological factors on wind power output was fully considered, and the Spearman correlation coefficient was used to select the meteorological factors most correlated with wind power output as model input parameters. Next, the Gaussian mixture model (GMM) was used to conduct cluster analysis on the wind power data. The elbow method was used to determine the optimal number of clusters, and in combination with the feature similarity and cosine similarity entropy methods, the most relevant cluster type for the day under test was determined among the historical data. Finally, the CNN-BiLSTM-Attention model was used for training to deeply extract the temporal features of wind power and obtain more accurate wind power prediction results. Taking the actual wind power data of Xinjiang region as an example, a simulation analysis was conducted. The verification results show that the prediction accuracy of the method is high, and it can provide strong support for the planning and stable operation of the power system.