Few-shot wind power prediction using sample transfer and imbalanced evolved neural network

Abstract

Accurate wind power prediction for newly built wind farms (NWFs) with limited historical data remains a significant challenge. To address this, we propose SDM-VMD-IENN, a novel framework integrating Similar Data Matching (SDM), Variational Mode Decomposition (VMD), and an Imbalanced Evolved Neural Network (IENN). This model uniquely combines data enhancement and evolutionary optimization to overcome the limitations of existing methods, including negative transfer effects in transfer learning models, data redundancy, and local convergence. Specifically, SDM mitigates negative transfer by filtering highly similar source domain data and constructing Gram matrix-based feature representations, enabling precise selection of high-similarity samples from the source domain. VMD decomposes non-stationary wind power sequences into stable subcomponents, reducing the nonlinear complexity of temporal features. IENN balances sample distribution discrepancies through evolutionary multi-loss optimization and adaptive weighting strategies based on distribution similarity, achieving global convergence. Experiments on real-world wind farms demonstrate that the proposed model exhibits higher prediction accuracy and enhanced robustness compared to classical models and other evolutionary frameworks, particularly under data scarcity scenarios. In our single-step and multi-step prediction tasks, SDM-VMD-IENN consistently outperforms traditional deep learning and evolutionary models. It effectively lowers RMSE and MAE. It is worth noting that in multiple experiments in case three, the SDM-VMD-IENN has a model that is superior to the single loss function. It highlights its strong generalization ability and applicability to data-scarce wind power prediction scenarios.