基于特征迁移学习和VMD-SE的超短期风电功率预测

Ultra-Short-Term Wind Power Forecasting Based on Feature Transfer Learning and VMD-SE

摘要: 针对新建风电场缺乏丰富的历史数据导致功率预测精度不足的问题，提出一种基于特征迁移学习和数据分解重构的超短期功率预测方法，用于提升小样本条件下的预测精度。首先，利用变分模态分解对历史风电功率数据进行分解，并根据样本熵的计算结果进行分组重构，结合风速、风向气象数据构建输入特征；其次，以数据丰富的源域风电场数据为基础，采用BiLSTM-CNN作为基础预测模型，结合最大平均差异和相关对齐作为新的分布差异度量，最小化源域与目标域之间的特征分布差异，同时引入对抗训练思想，利用域分类器作为对抗组件区分源域和目标域特征，促使特征提取器生成域不变的特征，从而实现源域风电场知识向目标域的有效迁移。实验结果表明，所提出的方法在数据稀缺的情况下能够有效提高风电场功率预测的准确性，为新建风电场的功率预测提供了新思路。

Abstract: To address the issue of insufficient power prediction accuracy in newly constructed wind farms due to limited historical data, this study proposes an ultra-short-term power prediction method based on feature transfer learning and data decomposition reconstruction to enhance prediction accuracy under small-sample conditions. First, historical wind power data is decomposed using variational modal decomposition and reconstructed into groups based on sample entropy calculations. These groups are then combined with meteorological data on wind speed and direction to form input features. Second, leveraging data-rich source domain wind farm data, a BiLSTM-CNN is employed as the base prediction model. This is integrated with maximum mean difference and correlation alignment as novel distribution divergence metrics, minimizing feature distribution differences between source and target domains. Concurrently, adversarial training principles are introduced, employing a domain classifier as the adversarial component to distinguish source-domain from target-domain features. This compels the feature extractor to generate domain-invariant features, thereby enabling effective knowledge transfer of source-domain wind farm characteristics to the target domain. Experimental results demonstrate that the proposed method significantly improves wind farm power prediction accuracy under data-scarce conditions, offering novel insights for power forecasting in newly constructed wind farms.