Multi-Step Wind Power Forecasting via CEEMDAN-Based Adaptive EWT and Frequency-Aligned GRU

Accurate multi-step wind power forecasting is important for power-system scheduling and renewable energy integration, yet the strong nonstationarity and multi-scale structure of wind power remain challenging. This study evaluates a CEEMDAN-based empirical wavelet transform (EWT) and frequency-aligned gated recurrent unit (GRU) framework for direct six-step forecasting. Each monthly series is first partitioned chronologically into training, validation, and test sets at a 7:1:2 ratio; CEEMDAN and EWT are then applied independently within the fixed partitions used for comparative evaluation. The decomposition atoms are aggregated into fixed physical frequency bands, providing a common representation across partitions. The validation set selects the EWT band subset used to denoise the leading high-frequency representation, and independent single-channel GRUs with temporal attention forecast the retained frequency-aligned components. Component forecasts are summed and constrained to the physical operating range. On twelve monthly subsets of the T1 SCADA dataset, the block-wise framework achieves equal-month averages of 3.9846% NMAE, 236.8220 kW RMSE, and 143.4440 kW MAE, outperforming a persistence baseline and seven trained benchmarks. The paired one-sided Wilcoxon tests show significant improvements over both CEEMDAN-GRU and CEEMDAN-EWT-LSTM for all three metrics. A strict origin-wise implementation, AdaptiveEWT+SDG, recomputes features only from the history available at each forecast origin. Across six initializations, it achieves 6.0050% NMAE, 376.2803 kW mean-horizon RMSE, and 216.1808 kW MAE, versus 6.0149%, 376.9031 kW, and 216.5353 kW for Persistence, confirming the adaptive-frequency principle under strictly sequential information.