Enhanced Dhole Optimization Algorithm for hyperparameter tuning of a TCN-BiGRU-MHA hybrid architecture for wind power forecasting

This study presents Temporal Convolutional Network-Bidirectional Gated Recurrent Unit-Multi Head Attention (TCN-BiGRU-MHA) hybrid deep learning model optimized with the Enhanced Dhole Optimization Algorithm (EDOA) for wind power estimation using real-world wind turbine data. The original Dhole Optimization Algorithm (DOA) exhibits limitations such as early convergence, getting stuck in local optima, and rapid population diversity reduction. EDOA addresses these limitations with three strategic improvements: (1) quantum-inspired mutation and differential evolution integration, (2) fitness and diversity-based enhanced spiral search, and (3) elite archive and multilayer recovery mechanism. In CEC 2019 benchmark tests, EDOA demonstrated superior performance over DOA in multimodal functions. The proposed TCN-BiGRU-MHA architecture systematically integrates multi-scale temporal feature extraction, bidirectional contextual learning, and dynamic feature weighting within a unified forecasting framework. Using real operational Supervisory Control and Data Acquisition (SCADA) data from the Nordex N117/3600 wind turbine in Yalova, Turkey and NASA MERRA-2 meteorological data, the EDOA-optimized model achieved Mean Absolute Error (MAE) = 76.26 kW, Root Mean Square Error (RMSE) = 137.47 kW, Coefficient of Determination (R 2 ) = 0.9888, and Weighted Absolute Percentage Error (WAPE) = 5.93%, surpassing all literature studies using the same dataset. With its excellent generalization capability and low overfitting tendency, the model offers reliable potential for use in real-time energy management systems.