Due to the high sensitivity of wind power generation to weather changes, the output power of wind farms fluctuates significantly over time. Traditional prediction models struggle to accurately forecast wind farm output power and effectively correct wind power forecasting errors. To address these issues, a PatchTST-POA-VMD-iTransformer hybrid prediction model is proposed. First, the Spearman’s rank correlation coefficient method is employed for quantifying the correlation between weather features and wind power, enabling data screening and preprocessing. Then, PatchTST is introduced for preliminary prediction of wind farm output power, yielding initial power forecasting results. Subsequently, pelican optimization algorithm (POA) optimized variational mode decomposition (VMD) is used to decompose wind power forecasting error sequence, and iTransformer is applied to further predict the decomposed error sequence. Finally, the preliminary power forecasting results are combined with the predicted error sequence to obtain the final wind power forecasting results. Ablation and comparative experiment results demonstrate that the proposed model achieves lower prediction errors and superior generalization ability, effectively improving the accuracy and reliability of ultra-short-term wind power forecasting.