基于气象场景条件风险最小化的多源迁移风速预测方法

Multi-Source Transfer Learning for Wind Speed Forecasting Based on Meteorological Scenario-Conditioned Risk Minimization

摘要: 针对新建风电场历史样本不足、跨场站分布差异及多源知识迁移易产生负迁移的问题，提出一种基于气象场景条件风险最小化的多源迁移风速预测方法。综合源域经验风险、条件分布差异、预测不确定性与梯度冲突，构建气象场景驱动的动态源选择机制；结合共享专属特征解耦、迁移门控与轻量残差适配，实现可迁移知识筛选与目标场站少样本适应。基于6个风电场开展留一场站实验，结果表明，本文方法在15 min和60 min预测任务中的MAE分别为0.448 m/s和0.972 m/s，较目标域独立训练的平均迁移增益分别达到37.34%和37.00%，负迁移率均为2.50%，验证了所提方法在少样本风速预测中的有效性。

Abstract: To address the challenges of insufficient historical samples at newly established wind farms, cross-site distribution discrepancies, and negative transfer arising from multi-source knowledge transfer, a multi-source transfer learning method for wind speed forecasting based on meteorological scenario-conditioned risk minimization is proposed. By jointly considering source-domain empirical risk, conditional distribution discrepancy, prediction uncertainty, and gradient conflict, a meteorological scenario-driven dynamic source selection mechanism is developed. Shared-private feature disentanglement, transfer gating, and lightweight residual adaptation are further integrated to enable transferable knowledge selection and few-sample adaptation at the target wind farm. Leave-one-site-out experiments are conducted on six wind farms. The results show that the proposed method achieves MAEs of 0.448 m/s and 0.972 m/s for the 15-min and 60-min forecasting tasks, respectively, with average transfer gains of 37.34% and 37.00% over target-domain independent training. The negative transfer rate is 2.50% for both forecasting horizons, demonstrating the effectiveness of the proposed method for few-sample wind speed forecasting.