Accurate wind power forecasting remains challenging due to the instability and volatility of wind power generation. Decomposition methods are widely used to improve forecasting performance by extracting complex fluctuation patterns from wind power series. However, previous decomposition-based models ignore the global interactions across sub-signals when forecasting the sub-signals separately and miss critical local details in sub-signals when modelling all the sub-signals in a single forecasting model. To address this issue, we propose a novel “Reconstruction-based Secondary Decomposition-Ensemble (RSDE)” framework for wind power forecasting, which simultaneously preserves the global interactions and local details. Firstly, an RSD method is adopted to extract fluctuation patterns from different frequency domains: decomposing the wind power by complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), reconstructing the primary sub-signals from specific frequency domains by spectral clustering (SC) and further decomposing the reconstructed sub-signals by singular spectrum analysis (SSA). Secondly, temporal convolutional networks (TCNs) are applied to forecast each reconstructed sub-signal by fitting the corresponding secondary decomposed sub-signals, which could effectively capture the local and global features from specific frequency domains. Finally, an ensemble strategy with error correction is adopted to obtain the final forecasting results by combining the forecasted reconstructed sub-signals and corresponding error forecasting results. Four wind power datasets with different time resolutions are introduced to evaluate the forecasting performance. The experimental results demonstrate that the proposed RSDE framework consistently outperforms the benchmark models. Moreover, the proposed sub-signal modelling strategy improves the forecasting performance by more than 30% on average, and the ensemble strategy with error correction improves the forecasting performance by about 10% on average.