Frequency-aware ultra-short-term wind power forecasting using CEEMDAN–VMD–SE and Transformer–GRU networks

Reliable short-term wind power forecasting is crucial for the effective management of renewable energy systems, as it enhances power grid scheduling and ensures system stability. Nonetheless, the sporadic and extremely unpredictable characteristics of wind power render the attainment of high forecasting accuracy a considerable challenge. The present study proposes a hybrid forecasting system that combines Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Variational Mode Decomposition (VMD), Sample Entropy (SE), Transformer, and Gated Recurrent Unit (GRU) to address this issue. The suggested method seeks to enhance the precision and resilience of ultra-short-term wind power forecasting by employing frequency characteristics decomposition and classification modeling. First, CEEMDAN is employed for decomposing the original wind power series into numerous Intrinsic Mode Functions (IMFs). The high-frequency components undergo additional denoising via VMD to mitigate noise interference. Then, utilizing the sample entropy values, the decomposed series are categorized into high-frequency and low-frequency components. Transformer and GRU models are respectively applied to predict these reconstructed sub-series. At the end of the predictions of all subseries are consolidated to get the ultimate wind power projection. Experimental results utilizing actual data from a wind farm in France confirm the enhanced forecasting precision of the proposed model, while also illustrating its robust generalization capability and practical applicability in real-world scenarios.