To improve the operational efficiency of wind farms and ensure the stable operation of the power system, an ultra-short-term wind power forecasting method based on information recombination and TCN-LSTM-MHSA is proposed. Using variational mode decomposition to split the power data, the subsequences are reorganized into high entropy sequences, medium entropy sequences, and low entropy sequences with different information types based on the evaluation results of sample entropy; Extracting feature representations of data through temporal convolutional network (TCN), further processing data features using the long short-term memory network (LSTM), and parallel learning of different attention representations using multi-head self-attention mechanism (MHSA) to construct a wind power prediction model. Using two datasets with different output capacity and fan composition as benchmarks to validate the model, the results show that the model in this paper has better prediction ability.