考虑样本加权的迁移学习暂态稳定评估模型更新方法

Transfer learning⁃based updating method of transient stability assessment model considering sample⁃weighted

在电力系统暂态稳定评估模型的更新过程中,针对与潜在故障相关性较小的故障样本影响迁移效果的问题,本研究从原始样本的特征量出发,发现其分布差异能反映故障之间的相关程度,由此提出考虑样本加权的迁移学习方法,进一步提高更新后评估模型的性能.首先,通过预先训练获得一个独立的域判别器,以此衡量训练模型的各故障样本相对于潜在故障的相似程度.其次,将量化后的分布差异通过密度比估计的方式进行转化,得到训练模型的各故障样本所赋予的权重大小.最后,将权重引入迁移学习更新评估模型的损失函数中,实现样本筛选.所提方法的有效性在IEEE-39节点系统和华东某区域的实际系统中均得到验证.

In the updating process of transient stability assessment model of power systems,aiming at the problem that the fault samples with less correlation to potential faults affect the migration effect,starting from the features of original samples,it is found that their distribution difference can reflect the degree of correlation between different faults.Therefore,a sample⁃weighted transfer learning method is proposed to further improve the performance of updated evaluation models.Firstly,an independent domain discriminator is obtained by pre⁃training,which can measure the similarity of each fault for model training with respect to potential faults.Then,the quantified distribution difference can be transformed by density ratio estimation to obtain the proper weights.Finally,these weights are introduced into the loss function of transfer learning to update evaluation models and achieve the purpose of sample screening.The effectiveness of the proposed method has been demonstrated in the IEEE 39⁃bus system and a realistic power system in East China.