A novel ultra-short-term wind power forecasting method based on TCN and Informer models

Accurate wind power forecasting can enhance the stability and reliability of the power grid, which is crucial for improving energy utilization efficiency and advancing the development of clean energy. A dual-channel network architecture, based on the integration of Temporal Convolutional Network (TCN) and Informer, is proposed in this study to improve the accuracy of wind power forecasting. Unlike traditional Variational Mode Decomposition (VMD) techniques, the wind power sequence is first decomposed into multiple intrinsic mode functions (IMFs) using VMD. Subsequently, sample entropy is innovatively utilized to classify the decomposed IMFs into two groups of differing complexity. Following this, both the emerging Informer and TCN models are concurrently applied to wind power forecasting, constituting the core methodology of this research. Based on the operational mechanisms of the two models, two integrated model architectures, Informer and TCN-Informer, are designed for the two groups of modes to extract feature information from the low-complexity and high-complexity mode groups, respectively. Finally, the prediction results of each IMF are aggregated and reconstructed to obtain the final forecast. We utilized a large-scale wind power dataset to train and test the proposed models, conducting predictions 15 min and 1 h in advance. The applicability of the models across different datasets was also validated. Experimental results demonstrate that, compared to commonly used wind power forecasting methods, the dual-channel architecture based on TCN and Informer models exhibits higher prediction accuracy and performance.