基于深度置信网络的电力系统暂态稳定评估方法  

A Method for Power System Transient Stability Assessment Based on Deep Belief Networks

作　　者：朱乔木 党杰[2] 陈金富[1] 徐友平[2] 李银红[1] 段献忠[1] 

机构地区：[1]强电磁工程与新技术国家重点实验室(华中科技大学):湖北省武汉市430074 [2]华中电网有限公司,湖北省武汉市430077

出　　处：《中国电机工程学报》2018年第3期735-743,共9页Proceedings of the CSEE

摘　　要：为进一步提高电力系统暂态稳定评估（transient stability assessment,TSA）的准确率,将＂深度学习＂方法引入电力系统暂态稳定评估,提出一种基于深度置信网络（deep belief networks,DBN）的TSA方法。构建一组能够反映系统暂态稳定特性的32维原始特征作为DBN模型的输入量,稳定结果作为输出量,利用深层架构对稳定结果与系统特征之间的映射关系进行训练。采用一种先使用无标注样本进行贪心无监督学习,后使用有标注样本进行监督学习的方法训练DBN模型的参数。训练后的模型能充分利用深层架构的特征提取优势,并能够利用大量无标注数据提高模型的泛化能力。新英格兰10机39节点系统上的仿真结果表明所提方法比常用的暂态稳定评估方法准确率更高,且能够在仅少量训练样本和含有无关特征的情况下获得优越的评估性能。To further improve the assessment accuracy of transient stability assessment（TSA） of power system, deep learning method was introduced into TSA for the first time, and a novel method for TSA based on deep belief networks（DBN） was proposed. A group of 32-dimensional original features reflecting status of power system＇s transient stability were employed as input, and the stability results were used as output of DBN model. The parameters of DBN model were firstly initialized by unsupervised learning method with no-label samples, and then fine-tuned by supervised learning method with labeled samples. The model is able to take the advantage of feature extraction of deep structure, and can enhance its generation ability by a large amount of no-label samples. Experiment results on New England 39-bus system demonstrate that the proposed method is more accurate than common TSA methods, and that the method performs very well with insufficient training samples or redundant features.

关 键 词：电力系统 暂态稳定评估 深度学习 深度置信网络 机器学习 

分 类 号：TM712[电气工程—电力系统及自动化]