Gated recurrent neural network enhanced wind power prediction accuracy using TPE Bayesian optimization

Accurate ultra-short-term (15-minute) wind power forecasting is essential for maintaining grid stability under high renewable penetration. This study proposes a systematic hyperparameter optimization framework based on the Tree-structured Parzen Estimator (TPE) for single-, double-, and triple-layer gated recurrent neural network architectures, including LSTM, GRU, and hybrid LSTM–GRU and GRU–LSTM models. Each architecture is allocated an equal search budget across all optimization methods, ensuring a fair comparison with conventional grid and random search. Experimental results demonstrate that TPE achieves faster convergence and lower optimization loss values, highlighting its efficiency in navigating complex hyperparameter spaces. Notably, a counterintuitive finding is observed: the optimized single-layer LSTM model outperforms all deeper architectures. Statistical validation using the Ljung–Box Q-test confirms that the residuals approximate white noise, indicating model adequacy and reliability. The proposed TPE-LSTM method achieves lower MAPE values compared to previous wind power prediction models.