In order to address the issues of underutilization of historical wind power information and insufficient exploration of the potential of machine learning models, a method for ultra-short-term wind power forecasting has been proposed. This method is based on feature singularity spectrum analysis and model error compensation. Firstly, random forest is used to analyze the influence of different features on the output power, and the cumulative contribution rate is used to extract the features. Secondly, by improving the Cao algorithm, the optimal embedding dimension for singular spectrum analysis is determined. The extracted features are denoised and a wind power prediction model is constructed based on the denoised data. Finally, the error prediction model is constructed by using the error between the predicted value and the real value, and the result of power prediction is corrected by the predicted error. The results from a small wind farm in China confirm that the proposed method reduced RSME and MSE by 45% and 53% compared to CNN-LSTM, thus verifying its effectiveness.