This study focuses on the importance of reliable surface wind forecasts for various sectors, particularly energy production. Traditional numerical weather prediction models are facing limitations and increasing complexity, leading to the development of machine learning models as alternatives or supplements. The research consists of two stages. In the first stage, the ERA5 database is used to evaluate the long-term performance of different combinations of features and two tree-based algorithms for predicting surface wind characteristics (speed and direction) in Cairo. The XGBoost algorithm slightly outperforms the Random Forest algorithm, especially when combined with appropriate feature selection. Even three years after the training period, the results remain very good, with an RMSE of 0.59 m/s, rRMSE of 17%, and R 2 of 0.84. The second stage assesses the multivariate approach's ability to forecast wind speed evolution at different time horizons (1–12 h) during a week characterized by significant wind dynamics. The forecasts demonstrate excellent agreement with observations at a 1-h time horizon, with an RMSE of 0.35 m/s, rRMSE of 7.6%, and R 2 of 0.98, surpassing or comparable to other literature results. However, as the time lag increases, the RMSE (0.86, 1.14, and 1.51 m/s for 3, 6, and 12 h, respectively) and rRMSE (18.7%, 24.8%, and 32.9% for 3, 6, and 12 h, respectively) also increase, while R 2 decreases (0.86, 0.79, and 0.60). Furthermore, the wind variations' amplitude is underestimated. To address this bias, a simple correction method is proposed.