钟光耀¹, 邰能灵¹(), 黄文焘¹, 李然¹, 傅晓飞², 纪坤华²

1.上海交通大学 电子信息与电气工程学院,上海 200240
2.国网上海电力公司,上海 200122

收稿日期:2021-06-11出版日期:2021-12-28发布日期:2021-12-30
通讯作者:邰能灵E-mail:nltai@sjtu.edu.cn
作者简介:钟光耀(1995-),男,浙江省宁波市人,硕士生,从事大数据在智能电网中的应用研究.
基金资助:中国南方电网有限责任公司科技项目(080037KK52180050GZHKJXM20180068);上海市教委科研创新重大项目(2019-01-07-00-02-E00044);国家重点研发计划(2019YFE0102900)

Attention Short-Term Forecasting Method of Distribution Load Based on Multi-Dimensional Clustering

ZHONG Guangyao¹, TAI Nengling¹(), HUANG Wentao¹, LI Ran¹, FU Xiaofei², JI Kunhua²

1. School of Electrical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
2. State Grid Shanghai Municipal Electric Power Company, Shanghai 200122, China

Received:2021-06-11Online:2021-12-28Published:2021-12-30
Contact:TAI Nengling E-mail:nltai@sjtu.edu.cn

摘要/Abstract

摘要： 在大规模配变负荷预测中,由于负荷特性差别以及受影响因素不同,若使用统一模型,准确率低且泛化能力差,若针对单台配变进行负荷预测建模,计算资源消耗过大.提出了一种基于多维聚类的配变负荷注意力长短期记忆网络(Attention Long Short-Term Memory,Attention-LSTM)短期预测方法.首先提取每个配变日负荷特征序列并利用非参数核方法进行概率拟合,形成配变负荷的典型日负荷序列;以欧式归整距离以及影响因素相似性作为相似度评判标准,使用改进的k均值聚类(k-means)双层聚类对日典型负荷序列进行负荷聚类分析;利用近邻传播(Affinity Propagation,AP)聚类提取影响因素相似时间序列,构建训练集,训练Attention-LSTM模型;针对不同的配变负荷类型以及不同的相似时间序列得到不同的Attention-LSTM模型.通过选取某市级配电网实测负荷数据以及气象等影响因素数据,验证了所提方法的有效性和实用性,准确率提升了2.75%且效率提升了616.8%.
关键词: 短期负荷预测, 日负荷序列, 负荷聚类, 相似时间序列, 长短期记忆网络
Abstract: Due to the difference in load characteristics and influencing factors in large-scale distribution transformer load forecasting, if all the distribution transformers share a unified model, the prediction accuracy is low, and if the model is built for each distribution transformer, the computational resources will be excessively consumed. An Attention-LSTM short-term forecasting method of distribution load based on multi-dimensional clustering is proposed. The non-parametric kernel method is used to perform probability fitting on the daily load characteristics to form a typical daily load sequence. Improved two-level clustering is applied for load clustering, taking the Euclidean warping distance and influence factors as the similarity evaluation criteria. AP clustering is utilized for obtaining similar time-series, and training sets are formed to train the Attention-LSTM model. Different Attention-LSTM models are obtained by training for different distribution load types and time-series. The effectiveness and practicability of the method proposed are verified by the load data and meteorological data of a municipal distribution network. The accuracy rate is increased by 2.75% and the efficiency is increased by 616.8%.
Key words: short-term load forecasting, daily load sequence, load clustering, similar time-series, long short-term memory network


PDF全文下载地址:

点我下载PDF