ML-Based Multi-Horizon Wind Speed and Wind Direction Forecasting for Aviation and Energy Applications in Coastal Crete

Accurate short-term and medium-term wind speed and wind direction forecasting are crucial for aviation safety, renewable energy management, and environmental planning, particularly in coastal areas with complex terrains. In this study, four machine learning models (Temporal Fusion Transformer (TFT), LightGBM, CatBoost, and a multi-head Convolutional Neural Network (CNN)) were used for multi-horizon (1–24 h) forecasting in Kastelli, Crete, near the new Heraklion International Airport, using a high-resolution multivariate meteorological dataset (2015–2023). For the wind speed forecasting, the best mean absolute error (MAE) values at each horizon are 1 h = 1.89 m/s (LightGBM), 6 h = 3.12 m/s (CatBoost), 12 h = 3.44 m/s (TFT), and 24 h = 3.38 m/s (TFT). For the wind direction forecasting, the best angular MAE values are 1 h = 8.66° (CatBoost), 6 h = 30.71° (CNN), 12 h = 35.29° (TFT), and 24 h = 25.65° (TFT). Overall, the study indicates that different models outperform at different horizons, with the tree-based models being the most effective for short-term forecasts, the convolutional network performing best at intermediate horizons, and the transformer-based architecture offering stronger performance over longer horizons. Compared to recent literature, the proposed framework achieves measurable improvements and confirms the feasibility of deploying ML-based forecasting tools.