Accurate short-term wind speed forecasting is crucial for efficient grid management and the integration of renewable energy sources. This study introduces a hybrid forecasting model that combines enhanced variational mode decomposition (EVMD), discrete wavelet transform (DWT), and an encoder-only Transformer. The dataset is first preprocessed to handle missing values and outliers and then subjected to a two-step decomposition process: EVMD extracts the first nine intrinsic mode functions (IMFs), and DWT is applied to the last IMF from EVMD to further decompose high-frequency components and capture finer-scale variations. The encoder-only Transformer is then used to process and forecast all subseries obtained from both EVMD and DWT, excluding the final IMF from EVMD, and their outputs are aggregated to produce the final prediction. The model was evaluated using data from the Thai Hoa wind farm in Binh Thuan, Vietnam, and compared with six benchmarks (variational mode decomposition (VMD)-Transformer, VMD-long short-term memory (LSTM), Transformer, LSTM, and random forest (RF)) using mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and normalized RMSE (NRMSE) for both single and multi-step forecasting. Results show that the proposed hybrid model consistently outperforms all benchmarks across all evaluation metrics, achieving an RMSE of 0.535 m/s for 1-step and 0.834 m/s for 24-step forecasting, representing up to 1.9% improvement over the best baseline and over 63% improvement compared to the worst-performing model, thereby demonstrating superior accuracy across both forecasting horizons. By leveraging multi-level decomposition and attention mechanisms, the model effectively captures complex wind speed patterns, enhancing forecasting performance and supporting better integration of wind energy into power systems.