Yan Mei, Jinxing Che, Qian Sun, Wei Dong
Accurate wind power forecasting is crucial for the effective scheduling and scientific management of wind energy, enhancing the safety and reliability of power grids. To handle the inherent intermittency of wind energy, forecasting methodologies have evolved from traditional statistical models to deep learning architectures. A dominant paradigm in the field has become the "decomposition-ensemble" framework, which disentangles complex wind power series into more predictable components. However, the efficacy of many existing models is constrained by key limitations, such as adopting a uniform sub-model for all components, insufficient validation on diverse data, and a general neglect of uncertainty quantification. To systematically address these challenges, this paper first provides a review of hybrid forecasting models. Subsequently, it presents a forecasting framework as a case study, which integrates multi-scale signal decomposition with an adaptive multi-model fusion strategy. This approach dynamically assigns suitable models to each decomposed component, and incorporates a non-parametric method for prediction interval generation. The framework’s performance was validated on two heterogeneous datasets representing different geographical locations, seasons, and temporal resolutions. The results demonstrate that the framework outperforms benchmarks in both accuracy and robustness. This study not only offers an effective solution but also highlights the importance of adaptive model selection and the quantification of forecasting uncertainty, charting a course for future research in reliable renewable energy forecasting. • Systematic review reveals gaps in decomposition-ensemble paradigms. • Case study validates an adaptive framework with dynamic model selection. • Framework robustness is verified on heterogeneous wind power datasets. • Future directions for reliable and risk-aware forecasting are outlined.