To improve the accuracy of wind power prediction, an ultra-short-term combination forecasting method based on Frequency Run Length Discriminant and Variational Modal Decomposition (VMD) residual error correction is proposed. Firstly, the original wind power sequence is decomposed by VMD to obtain a series of subsequences with different center frequencies, and then the residual sequence is extracted from the difference in the se? quences. The residual sequence has the characteristics of large fluctuation, nonlinear complexity, and unsteadiness, which inherits the original sequence noise component and the masked information during decomposition, and the adaptive t-distribution Sparrow Search Algorithm (t-SSA-LSTM) combined with the weather features is used for the prediction. The sub-sequences are divided into two kinds of signals class, namely high and low- frequency se? quences, by using the Frequency Run Length Discriminant method. The low-frequency sequences are linear stability and the adaptive t-distribution Sparrow Search Algorithm (t-SSA) is used to optimize the autoregressive integrated moving average (ARIMA) model prediction. The characteristics of high-frequency sequences are volatile and com? plex, and the t-SSA is used to optimize the Long Short-Term Memory (LSTM) neural network for the prediction of high-frequency sequences. Finally, the wind power prediction results are achieved by linearly superimposing the pre? diction results of different sequences. The proposed model is finally applied to a wind farm in China, and the results show that the model can effectively improve the prediction accuracy.