Adaptive wind power forecasting with multi-level decomposition and multi-branch attention fusion
Haitao Xiong, Ping Zhao, Yuanyuan CaiPurpose
Accurate wind power forecasting is essential for secure and economical power system operation, yet the strong non-stationarity and multiscale fluctuations of wind power signals pose significant challenges to conventional models.
Design/methodology/approach
This study proposes M2DAF, a hybrid framework that integrates an adaptive multilevel decomposition stage with a multi-branch prediction stage. The decomposition stage employs dynamic complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) with adaptive noise scaling, hierarchical clustering based on morphological features, and variational mode decomposition (VMD) refinement to generate well-separated and physically interpretable components. The prediction stage adopts a parallel convolutional neural network–long short-term memory–Transformer (CNN-LSTM-Transformer) network to capture local, mid-range, and long-range temporal dependencies. An adaptive attention fusion module dynamically adjusts branch contributions at each prediction step.
Findings
Extensive experiments on three wind farm datasets from France, Turkey, and Spain demonstrate that M2DAF consistently outperforms nine benchmark methods across multiple evaluation metrics. Ablation studies confirm that the decomposition modules contribute most significantly to the overall accuracy. Additional experiments on multi-step forecasting, multivariate inputs, and ramp event prediction under extreme conditions further validate the framework's robustness and cross-regional adaptability across diverse climatic regimes. Overall, the results highlight the effectiveness of the proposed multilevel decomposition and adaptive fusion strategy in capturing the complex, multiscale dynamics of wind power generation.
Originality/value
This study introduces an adaptive multilevel decomposition and multi-branch attention fusion framework, which innovatively integrates dynamic CEEMDAN, hierarchical clustering, VMD refinement, and a parallel CNN-LSTM-Transformer network with real-time adaptive fusion. This approach effectively addresses the non-stationarity and multiscale nature of wind power data, providing a robust, adaptable solution for highly accurate and reliable wind power forecasting across diverse climatic conditions.