Accurate wind speed forecasting plays a crucial role in supporting early warning systems for extreme wind events. However, the inherent non-linearity and non-stationarity of wind speed data pose significant challenges. This study addresses these issues by evaluating the effectiveness of targeted Empirical Wavelet Transform (EWT) denoising applied to specific Intrinsic Mode Functions (IMFs) derived from Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN). Daily wind speed data from 2000 to 2023 were decomposed using ICEEMDAN, and denoising was selectively applied to IMF1, IMF2, and IMF3. Each IMF was then modeled using a Bidirectional Long Short-Term Memory (BiLSTM) network under a time-series cross-validation framework. Among all model configurations, the ICEEMDAN+EWT(IMF1 & IMF2)+BiLSTM model achieved the highest predictive accuracy, with an R² of 0.8885, RMSE of 0.501, and MAPE of 7.64%. This result outperformed both the baseline BiLSTM model (R² = 0.0501) and the ICEEMDAN+BiLSTM model without EWT denoising (R² = 0.6433). Moreover, denoising on IMF1 alone also yielded a strong performance (R² = 0.8879), emphasizing the importance of early component selection. Conversely, applying EWT to IMF2 or IMF3 individually resulted in lower R² values of 0.6639 and 0.6327, respectively, indicating limited individual contribution. These findings confirm that selective denoising, especially on the high-frequency IMFs, substantially enhances forecasting accuracy. The proposed approach holds significant potential to improve the timeliness and reliability of wind-related early warning systems, thus contributing to more effective disaster risk reduction strategies.