Wind speed prediction using feature engineering–driven deep learning on long-term meteorological data: A multi-city and ablation-based analysis
Journal of Atmospheric and Solar-Terrestrial Physics, cilt.287, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 287
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.jastp.2026.106940
- Dergi Adı: Journal of Atmospheric and Solar-Terrestrial Physics
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Artic & Antarctic Regions, Compendex, INSPEC, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO)
- Anahtar Kelimeler: Ablation study, BiLSTM, CNN, Feature engineering, Hybrid DL, Wind speed forecasting
- Hakkari Üniversitesi Adresli: Evet
Özet
Accurate wind speed forecasting remains a challenging task because atmospheric processes exhibit highly nonlinear, stochastic and region-dependent characteristics. Rather than proposing another increasingly complex deep learning (DL) architecture, this study presents a feature engineering-driven forecasting framework that systematically integrates cyclic encoding, differential features and moving average smoothing with deep learning models to improve forecasting accuracy while maintaining computational efficiency. Five representative models, namely CNN, LSTM, stacked BiLSTM, FE-GRU and a CNN–BiLSTM–Linear stacking ensemble, were comparatively evaluated under identical experimental conditions using a 13-year (2010–2022) hourly meteorological dataset collected from three climatically distinct regions of Türkiye (Çanakkale, Niğde and Van), each containing more than 100,000 observations. Experimental results consistently demonstrate that the proposed FE-GRU model achieved the lowest relative prediction errors across all study regions when evaluated using Mean Absolute Percentage Error (MAPE), which was adopted as the primary performance criterion. The average MAPE values obtained for Çanakkale, Niğde and Van were 1.20%, 1.31% and 1.72%, respectively, while maintaining high goodness-of-fit (R2 ≥ 0.995) and low Root Mean Square Error (RMSE) values. Although the CNN–BiLSTM–Linear stacking ensemble produced competitive RMSE values, it required substantially higher computational cost, whereas FE-GRU achieved a superior balance between forecasting accuracy and computational efficiency. The reliability of the proposed framework was further supported through paired statistical significance tests, confidence interval analysis, Taylor diagrams and feature importance evaluation, demonstrating consistent agreement with observed wind speed measurements under diverse geographical and climatic conditions. Overall, the results indicate that explicit feature engineering contributes more effectively to forecasting performance than increasing architectural complexity alone. The proposed framework therefore provides an accurate, computationally efficient and geographically robust solution for practical wind speed forecasting and renewable energy management applications.