摘要
基于深度学习的暂态稳定评估模型通常需要大量的有标注样本用于离线训练。一旦电网的运行方式和拓扑结构发生较大变化,预训练模型的性能将劣化甚至失效,使得在线评估时存在一定的空窗期。为了解决这一问题,以深度置信网络(DBN)为研究载体,将深度学习、迁移学习和主动学习相结合,提出一种基于DBN模型的主动迁移学习方法。首先,训练DBN来挖掘输入特征和暂态稳定评估结果间的映射关系,得到更好的暂态稳定评估效果。其次,当拓扑结构和运行方式发生较大变化时,通过短期仿真生成大量的无标注样本,利用主动学习来筛选少量最富有信息的样本,并通过长期仿真对这部分样本进行标注,显著减少了样本的生成时间。最后,计算源域和目标域数据分布的最大均值差异,选择不同的迁移路径,在确保迁移效果的前提下进一步缩短了迁移时间。采用新英格兰10机39节点系统、NPCC 48机140节点系统和中国华中电网进行了仿真,结果验证了所提方法具有高精度、快速性和鲁棒性,有效缩短了深度学习模型在线应用时的空窗期。
Transient stability assessment models based on deep learning usually require a large number of labeled samples for offline training. Once the topologies or operation conditions of power grids change greatly, the pre-trained model deteriorates in performance and even becomes ineffective, resulting in a certain blank window during the online assessment. In order to solve this problem, taking deep belief network(DBN) as the research carrier, this paper combines deep learning, transfer learning and active learning to propose an active transfer learning method based on DBN model. Firstly, DBN is trained to mine the mapping relationship between input features and transient stability assessment results, so as to obtain better effect for transient stability assessment. Secondly, when the topologies or operation conditions change substantially, a large number of unlabeled samples are generated through short-term simulation, and active learning is exploited to select the minimum number of samples with the most information. Then, these selected samples are labeled by long-term simulation, which effectively reduces the time of sample generation. Finally, the maximum mean discrepancy(MMD) of the data distribution between the source domain and the target domain is calculated to select different transfer paths. The transfer time is further shortened on the premise of ensuring the transfer effect. The simulations on New England 10-unit 39-bus system, NPCC 48-unit 140-bus system and central China power grid are carried out, and the results verify that the proposed method has high accuracy, rapidity and robustness, which effectively shortens the blank window period of online application of the deep learning model.
作者
李宝琴
吴俊勇
李栌苏
史法顺
赵鹏杰
王燚
LI Baoqin;WU Junyong;LI Lusu;SHI Fashun;ZHAO Pengjie;WANG Yi(School of Electrical Engineering,Beijing Jiaotong University,Beijing 100044,China)
出处
《电力系统自动化》
EI
CSCD
北大核心
2023年第4期121-132,共12页
Automation of Electric Power Systems
基金
国家重点研发计划资助项目(2018YFB0904500)
国家电网有限公司科技项目(SGLNDK00KJJS1800236)。
关键词
深度学习
暂态稳定评估
迁移学习
主动学习
深度置信网络
deep learning
transient stability assessment
transfer learning
active learning
deep belief network