首页|期刊导航|浙江电力|基于双层聚类和时空图注意力网络的台区集群电压预测

基于双层聚类和时空图注意力网络的台区集群电压预测OA

Voltage prediction for distribution transformer clusters based on dual-layer cluster-ing and a spatiotemporal graph attention network

中文摘要英文摘要

当前配电变压器(以下简称"配变")的台区电压预测研究多以时间序列建模的单点预测为主,对配变间复杂的时空相关性建模不充分,大规模配变集群场景下的扩展性和计算效率存在不足.为此,提出一种基于长短时间尺度双层聚类与时空图注意力网络的电压预测方法.首先,在短时间尺度上刻画典型日的电压波动特征,构建运行特性相似的配变子集群;其次,在长时间尺度上提取年运行水平特征,形成双层聚类结构;在此基础上,构建融合电压、有功功率及地理位置三类异构图结构的多通道图注意力网络,实现对集群电压的精准预测.最后,基于上海市某区域配变的真实运行数据进行算例仿真,结果验证了该方法在提升配变电压预测精度方面的有效性.

Existing research on voltage prediction for distribution transformers(hereinafter referred to as"DTs")primarily relies on single-point time-series modeling,which insufficiently captures complex spatiotemporal correla-tions among DTs.This leads to limited scalability and computational efficiency in large-scale DT cluster scenarios.To address these issues,a voltage prediction method based on dual-layer clustering across long-and short-term time scales and a spatiotemporal graph attention network is proposed.First,on a short-term time scale,voltage fluctua-tion characteristics of typical daily profiles are extracted to construct sub-clusters of DTs with similar operating be-haviors.Second,on a long-term time scale,annual operational features are extracted to form a dual-layer clustering structure.On this basis,a multi-channel graph attention network is developed by integrating heterogeneous graph structures of voltage,active power,and geographical location,enabling accurate prediction of cluster-level voltage.Finally,case studies using real operational data from a distribution transformer in Shanghai validate the effective-ness of the proposed method in improving voltage prediction accuracy for distribution transformers.

陈以恒;杨秀;孙改平;杨帆

上海电力大学 电气工程学部,上海 200090上海电力大学 电气工程学部,上海 200090上海电力大学 电气工程学部,上海 200090国家电网上海电力科学研究院,上海 200437

配变电压预测多时间尺度特征双层聚类时空相关性图神经网络多通道注意力机制

distribution transformer voltage predictionmulti-time-scale featuresdual-layer clusteringspatiotem-poral correlationgraph neural networkmulti-channel attention mechanism

《浙江电力》 2026 (9)

79-89,11

国家自然科学基金(52177098)上海电力人工智能工程技术研究中心项目(19DZ2252800)

10.19585/j.zjdl.202609008

评论