解决风速预测中的少样本问题：一种基于分解与学习集成的新型迁移策略

Abstract With the continuous growth in demand for renewable energy , new wind power farms are being built globally. However, due to the limited availability of wind speed data for new equipment, directly forecasting wind speed data for new turbines has become extremely challenging. To address this issue, this paper proposes a rapid transfer strategy for the few-shot problem. The transfer framework is constructed in two stages. The first stage involves pre-training the model on large sample data. Initially, Dynamic Time Warping (DTW) is used to select datasets most similar to the target domain. Then, variational mode decomposition (VMD) is employed to divide the dataset into different modes, and the modal components with similar complexity are reassembled into new components based on sample entropy (SE) theory. For each new component, the most suitable deep learning structure is selected from a model selection pool using customized evaluation criteria. Finally, the learning ensemble method combines the prediction results of each model to obtain the final prediction for the source domain. Compared with traditional linear ensemble methods, the learning ensemble can capture nonlinear features , significantly improving prediction accuracy. The second stage involves transferring the pre-trained model to the target domain. Initially, the prediction models for each component are tested directly on the target domain, and based on the results, a decision is made on whether to fine-tune. If any models require fine-tuning, the learning ensemble must also be fine-tuned simultaneously. We simulate the entire transfer learning process using wind speed data from a wind farm in Queensland. Experimental results show that the proposed strategy effectively addresses the few-shot problem in wind speed prediction with minimal transfer cost. The robustness and generalization of the proposed strategy are verified by using four sets of data from different locations.