Transfer learning based-hybrid model for short-term wind speed forecasting

Reliable integration of wind energy into the power grid relies on accurate short-term wind speed forecasting. To address this challenge, this paper introduces a hybrid forecasting framework that combines Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Bidirectional Gated Recurrent Unit Encoder-Decoder (Bi-GRU-ED), and transfer learning. CEEMDAN decomposes the original wind speed series into intrinsic mode functions (IMFs) and a residual, enabling the capture of nonlinear characteristics while reducing noise interference. The forecasting process begins with forecasting the residual component using the Bi-GRU- ED Res model. Subsequently, transfer learning is employed to reduces computational demands and boosts performance by leveraging pre-trained knowledge across decomposition components to efficiently forecast the remaining IMFs, reducing training time without compromising accuracy. Finally, all forecasted components are aggregated to reconstruct the wind speed series. The proposed approach was validated on two datasets from the Maghreb region and compared with multiple benchmark models. Results demonstrate that the hybrid model achieves superior forecasting accuracy, confirming its suitability for short-term wind speed prediction and potential application in renewable energy systems. • Proposed a hybrid CEEMDAN-BiGRU-ED model for wind speed forecasting. • CEEMDAN decomposition minimizes noise and captures nonlinear patterns. • Bi-GRU Encoder-Decoder models temporal dependencies effectively. • Transfer learning boosts prediction accuracy and model adaptability. • The proposed model outperforms benchmarks across all evaluation metrics.