基于主动迁移学习的电力系统拓扑自适应暂态稳定评估

An Active Transfer Learning Scheme for Power System Transient Stability Assessment Adaptive to the Topological Variability

在基于深度学习的数据驱动型电力系统暂态稳定评估技术中,需要解决的一个关键应用挑战是对系统拓扑结构变化和电网扩展的适用性问题。该文首先提出一种考虑电气属性的互注意力图池化方法,基于注意力机制,按照电网节点的属性和距离执行差异化的特征综合,提升图深度学习特征聚合对拓扑结构变化的适用性。在此基础上,提出一种基于主动迁移学习的暂态稳定评估方案,在电网结构大幅变化或扩展时可以利用少量新系统标注样本完成模型的学习。提出梯度加权距离评价样本相似性,从原系统中获取新系统样本伪标签以实现新系统模型的初始训练。设计考虑不确定性和多样性的主动采样策略从新系统中持续挑选高价值样本进行标注,大幅减少样本标注成本。在IEEE 39、300节点系统上的测试结果验证了该文方案的优越性。

A key application challenge to deep learningbased data-driven power system transient stability assessment,is the adaptability of model to system topology changes and grid expansion.This paper first proposes a method called Electrical Attribute-based Interaction Attention Graph Pooling(EIAPool),aggregating power grid node features differently based on their attributes and distances,through attention mechanism,to improve the adaptability of graph deep learning to the topological variations.On this basis,a transient stability assessment scheme based on active transfer learning(ATL-TSA),requiring fewer labelled examples to complete model training in the new system,is proposed to against great variations of grid topology.ATL-TSA evaluates the similarity between instances by the gradient weighted distance,and obtains the pseudo-labels of new instances from the original system,to train a initial model in the new system.Active learning strategy considering uncertainty and diversity,queries the labels of the most useful samples from the new system and,thus,significantly saves the labeling cost.Experiments on the IEEE 39-bus and 300-bus system show our method and scheme to be superior.