摘要
“双碳”目标旨在推动能源转型与减排,新型电力系统作为关键,促进清洁能源接入与利用,减碳效果显著。但其多元化负荷结构增大了预测难度。为应对“双碳”要求,解决新型电力系统中多节点负荷预测的复杂时空依赖性和非线性问题,文章提出了一种基于多尺度自适应时空图卷积网络(adaptive spatio-temporal graph convolutional network,ASTGCN)与基于Transformer的双边编码器表示(bidirectional encoder representations from transformers,BERT)模型的多节点短期负荷预测方法。首先,采用Prophet算法对负荷数据进行拟合分解,获取不同尺度下的负荷数据分量,并与强相关的天气数据共同构建多元数据集;其次,引入可膨胀的滑动时空窗口和时空图卷积算子构建ASTGCN,同时捕捉空间和时间上的复杂依赖关系,并引入BERT模型对时间序列数据进行编码,利用其强大的处理能力来捕捉负荷数据中的长期依赖性;最后,用门控融合网络对两个模型进行融合。基于美国纽约州的公开数据集进行测试,单日和单周的测试结果均表明所提模型不仅能有效挖掘节点的耦合特性,还能补充挖掘中长期时序特征,并显著提升预测精度,降低预测误差。
The“dual-carbon”goal aims to promote energy transition and emission reduction,and the new power system,as a key to promoting clean energy access and utilization,has a significant carbon reduction effect.However,its diversified load structure increases the difficulty of forecasting.To cope with the requirements of“dual-carbon”and to solve the complex spatial and temporal dependence and nonlinear problems of multi-node load forecasting in the new power system,this paper presents a multi-node short-term load forecasting approach based on multi-scale adaptive spatio-temporal graph convolutional network(ASTGCN)and bidirectional encoder representations from transformers(BERT)model.First,the load data is fitted and broken down using the Prophet method.Then,the load data components are extracted at various scales,and a multivariate data set is created using the heavily correlated weather data.Second,ASTGCN is employed to capture complicated spatial and temporal dependencies simultaneously.The BERT model is presented to encode time series data to capture long-term relationships in load data with its strong processing capability.Ultimately,a gated fusion network is employed to merge the two models.Based on the public dataset of New York State in the U.S.,the test results of both single-day and single-week show that the model proposed in this paper can not only effectively mine the coupling characteristics of the nodes but also supplement the mining of the medium-long-term time-series features,and significantly improve the prediction accuracy and reduce the prediction error.
作者
吴兴扬
戴剑丰
WU Xingyang;DAI Jianfeng(Department of Mechanical&Electronic Engineering,Jinling Institute of Technology,Nanjing 211169,Jiangsu Province,China;College of Automation&College of Artificial Intelligence,Nanjing University of Posts and Telecommunications,Nanjing 210023,Jiangsu Province,China)
出处
《电网技术》
北大核心
2025年第9期3756-3766,I0072-I0075,共15页
Power System Technology
基金
国家自然科学基金项目(62173188)
金陵科技学院高层次人才科研启动基金(jit-b-202403)。
