Use of Advanced Optimization and Hybrid Deep Learning Models in Ultra-Short-Term Wind Energy Forecasting: The Case of Hatay, Türkiye

Authors : Orkun Teke, Tolga Depci

Pages : 214-229

Doi:10.59940/jismar.1787678

View : 57 | Download : 116

Publication Date : 2025-12-31

Article Type : Research Paper

Abstract :This study presents a performance comparison of hybrid deep learning approaches with an end- to-end data pipeline designed to enhance accuracy and stability in ultra-short-term wind power forecasting. Anomaly removal using DBSCAN and feature selection based on RFECV are applied to multivariate SCADA-based data. An advanced hyperparameter optimization tool, Optuna, trained the models (SDAE baseline, CNN-LSTM, and GRU-LSTM) using progressive search and pruning strategies. Performance is evaluated using MAE, RMSE, and R² metrics from t+1 up to t+6 horizons. The findings indicate a significant superiority of the hybrid architectures over the baseline (SDAE) model: CNN-LSTM maintains consistently high accuracy across all horizons, while GRU-LSTM yields the lowest error metrics specifically at the shortest horizon (achieving an R²=0.9976 at t+1). The stability of the CNN-LSTM is maintained as the forecasting horizon extends, achieving a respectable performance of R²=0.79 even at t+6. This work proposes the operational use of GRU-LSTM for the shortest-term forecasts and CNN-LSTM for more stable predictions as the horizon lengthens. The results demonstrate that hybrid models establish a reliable foundation for industrial applications and suggest further gains are possible through the integration of uncertainty modeling and Numerical Weather Prediction (NWP).
Keywords : Rüzgar gücü kestirimi, ultra-kısa dönem tahmin, CNN-LSTM, GRU-LSTM, RFECV, Optuna, DBSCAN, SCADA

ORIGINAL ARTICLE URL