Accurate wind power prediction is crucial for ensuring the stable operation of power grids and improving the efficiency of wind resource utilization. To address the non-stationary and intermittent characteristics of wind power data, this paper proposes a combined IZOA-Transformer-BiGRU prediction model based on data decomposition techniques to enhance the accuracy and reliability of short-term wind power forecasting. First, the energy difference method is employed to determine the number of sub-modalities for variational mode decomposition, which decomposes the original wind power with strong random fluctuations into a series of relatively stable sub-sequences, enabling better more effective extraction of temporal features. Next, the Transformer-BiGRU model is constructed, incorporating a multi-head attention mechanism to process interactions between multiple features in parallel, while the BiGRU component captures temporal dependencies within the sequence, thus enhancing prediction performance. To further improve the model’s forecasting accuracy, an improved zebra optimization algorithm, integrating singer chaotic mapping, lens refraction-based learning, and the simplex method, is developed to optimize four key hyperparameters of the Transformer-BiGRU model: the number of hidden layer neurons, initial learning rate, regularization coefficient, and the number of attention heads. Finally, the IZOA-Transformer-BiGRU model predicts the subsequences derived from VMD, and the final prediction is reconstructed through aggregation. Experimental results show that, compared to the standalone BiGRU model, the proposed model improves the coefficient of determination by 5.10% and reduces the mean absolute error, root mean square error, and mean absolute percentage error by 56.17%, 54.58%, and 54.55%, respectively. demonstrating its high prediction accuracy.