A multi-channel spatiotemporal SegNet model for short term wind power prediction with sequence decomposition and reconstruction

• A multi-channel spatiotemporal SegNet (MCST-SegNet) model is proposed to achieve synchronous power prediction for all wind turbines in a wind farm. • The SegNet architecture is improved by incorporating temporal learning capabilities to adapt to the joint training of multi-dimensional spatiotemporal tensors. • A strategy using the Maximum Information Coefficient (MIC) to optimize Complementary Ensemble Empirical Mode Decomposition (CEEMD) is adopted to enhance information extraction efficiency. • The proposed method was applied to a wind farm with complex terrain and compared with other methods High precision wind power prediction can serve as an important reference for wind power grid integration. However, the complex spatiotemporal correlations between wind turbine operations and the high randomness of wind energy pose serious challenges to accurate wind power sequence prediction. This article proposes a multi-channel spatiotemporal SegNet (MCST-SegNet) model that achieves synchronous power prediction for all wind turbines in a wind farm. In the MCST-SegNet framework, each variable sequence matrix is first feature-expanded through a channel generation layer, converting the output into a single channel spatiotemporal feature map. Subsequently, all spatiotemporal feature maps are merged into a multi-channel spatiotemporal tensor, which enters the SegNet architecture capable of completing spatiotemporal joint training. The original SegNet network is improved into an encoder-predicting architecture with temporal learning capability to adapt to multi-dimensional spatiotemporal tensor joint training. In data processing, a sequence decomposition and reconstruction strategy using maximum information coefficient (MIC) to optimize complementary ensemble empirical mode decomposition (CEEMD) has been proposed. This strategy maximizes the extraction of information related to input variables and power without increasing input complexity. The experimental results show that the proposed method outperforms multiple classical and advanced benchmarks in spatiotemporal joint training.