Dynamic graph convolutional network considering wind speed delay and two-stage transfer learning applied to few-shot wind power prediction

To address the performance degradation of prediction models caused by limited data in newly built wind farms, this study proposes a time-varying physics-constrained "multi-source domain to pseudo to source domain - target domain" two-stage transfer learning framework by revealing the dynamic spatiotemporal coupling mechanism between adjacent wind farms induced by wind direction and speed propagation delays. First, the multi-kernel maximum mean discrepancy (MK-MMD) algorithm is introduced to capture both local features and global distribution similarities between source and target domains, enabling accurate selection of highly relevant source wind farms. Then, by combining geographical locations, wind direction, and wind speed, a bidirectional wind delay criterion is formulated to quantify the time-varying delays between wind farms. Based on this, constructed a wind direction and speed constrained dynamic graph convolutional network (WDSDGCN). WDSDGCN projects historical source domain data onto the target domain coordinates to generate pseudo-source domain samples that are structurally consistent with the target. Finally, a temporal convolutional network (TCN) is pre-trained on the pseudo-source domain and fine-tuned with limited target data. Experimental results on four datasets demonstrate that the proposed method consistently achieves R2 scores above 91.07 %, showing excellent accuracy and robustness in few-shot wind power prediction.