摘要
配变台区低压跳闸预测对供电部门掌握配电网运行态势至关重要。针对传统处理样本不平衡抽样方法容易造成信息丢失、引入噪声的缺陷,提出了一种领域清理(neighbourhood clean rule,NCL)欠采样和生成对抗网络组合的不平衡数据处理方法。首先采用NCL欠采样合理清除正常运行样本;然后采用稀疏自编码器学习大规模配变影响因素的内在特征,使用生成对抗网络去拟合、生成新的特征表示,并对其解码得到新跳闸样本;最后为了处理高维特征输入下的分类问题,使用随机森林分类器对目标配变进行低压跳闸预测,并依据跳闸概率进一步建立风险等级划分机制。以某地区配变为例进行建模预测,实验结果表明,所提模型在目标配变的预测准确率为99.65%,能有效预测低压跳闸事件和定位高风险台区。
It is vital for the power supply department to master the operation situation of distribution network to predict the low-voltage trip of distribution transformer area.Aiming at the defects of traditional unbalanced sampling methods,such as information loss and noise,we proposed an unbalanced data processing method combined with neighbourhood clean rule(NCL)under-sampling and generating counter network.Firstly,the NCL under-sampling was used to remove the normal operation samples reasonably.Then sparse auto-encoder was used to learn the internal characteristics of the influencing factors of large-scale distribution transformer,and generative adversarial network was used to fit and generate new feature representation,and new tripping samples were obtained by decoding.In order to solve the classification problem under high-dimensional feature input,a random forest classifier was used to predict the low-voltage trip of the target distribution transformer,and a risk classification mechanism was established based on the trip probability.A regional distribution transformer is taken as an example to establish a model for prediction,and the experimental results show that the accuracy of the proposed model in the target distribution transformer is 99.65%,which can effectively predict the low-voltage trip events and locate the high-risk station area.
作者
殷豪
丁伟锋
陈嘉铭
陈顺
欧祖宏
孟安波
YIN Hao;DING Weifeng;CHEN Jiaming;CHEN Shun;OU Zuhong;MENG Anbo(College of Automation,Guangdong University of Technology,Guangzhou 510006,China)
出处
《高电压技术》
EI
CAS
CSCD
北大核心
2021年第7期2321-2329,共9页
High Voltage Engineering
关键词
配变台区
低压跳闸预测
NCL欠采样
稀疏自编码器
生成对抗网络
随机森林
distribution transformation area
low-voltage trip prediction
NCL under-sampling
sparse auto-encoder
generative adversarial network
random forest
