基于分层有向图与动态时空相关性的小区域光伏超短期预测方法  

Small Area PV Ultra-short-term Prediction Method Using Stratified Digraph and Dynamic Spatial-temporal Correlation

作　　者：欧阳永健 缪希仁[1] 林蔚青 黄燕帼 OUYANG Yongjian;MIAO Xiren;LIN Weiqing;HUANG Yanguo(College of Electrical Engineering and Automation,Fuzhou University,Fuzhou 350108,Fujian Province,China;Longyan Power Supply Company,State Grid Fujian Electric Power Co.,Ltd.,Longyan 364000,Fujian Province,China)

机构地区：[1]福州大学电气工程与自动化学院,福建省福州市350108 [2]国网福建省电力有限公司龙岩供电公司,福建省龙岩市364000

出　　处：《电网技术》2024年第6期2458-2468,I0055,共12页Power System Technology

基　　金：福建省高校产学合作项目(2023H6006);国网福建省电力有限公司项目(521360230001DX)。

摘　　要：小区域光伏发电功率准确预测,已成为高渗透多台区系统精准调控运行的技术瓶颈。目前光伏发电功率的预测方法缺乏对台区小区域内光伏集群效应的考虑,忽略了输入变量的内在因果关系以及动态相关性。针对于上述问题,提出一种基于分层有向图以及动态图卷积循环网络(dynamic graph convolutional recurrent network,DGCRN)的小区域光伏预测方法。首先,考虑出力数据与数值天气预报(numerical weather prediction,NWP)单向关系,生成具有因果联系的分层有向图。其次,依据节点属性在每个时间步生成动态图,与预先定义的静态图有机结合,以捕捉节点之间的动态时空相关性。最后,将具有动态时空相关性的图结构用于模型训练。以某个小区域台区29个光伏节点加以预测建模,实验结果表明,DGCRN模型具备捕获多参量间的因果规律和提取光伏功率的短期动态特征的能力,其在小区域多节点的光伏发电功率预测性能优越。The accurate prediction of photovoltaic power generation in small areas has become a technical bottleneck for the precise control and operation of high-penetration multi-zone systems.Currently,photovoltaic power generation prediction methods need more consideration of the photovoltaic cluster effect in small areas of the station area and ignore the inherent causality and dynamic correlation of input variables.To solve these problems,a small area PV prediction method based on a hierarchical digraph and dynamic graph convolutional recurrent network(DGCRN)is proposed.Firstly,considering the unidirectional relationship between output data and numerical weather prediction(NWP),a hierarchical digraph with a causal relationship is generated.Secondly,a dynamic graph is generated at each time step according to node attributes,which is organically combined with a pre-defined static graph to capture the dynamic spatio-temporal correlation between nodes.Finally,the graph structure with dynamic spatiotemporal correlation is used for model training.The experimental results show that the DGCRN model can capture the causal law between multiple parameters,extract the short-term dynamic characteristics of photovoltaic power,and have superior performance in predicting photovoltaic power in a small area with multiple nodes.

关 键 词：分层有向图 动态相关性 图卷积网络 区域台区光伏预测 

分 类 号：TM615[电气工程—电力系统及自动化]