基于多神经网络混合模型的10 k V配电台区低电压中长期预测  

Low Voltage Medium to Long-Term Prediction in 10 kv Distribution Station Area Based on Multiple Model

作　　者：覃日升 张辰 马月 李颛 彭偲达 孙国芳 Qin Risheng;Zhang Chen;Ma Yue;Li Zhuan;Peng Sida;Sun Guofang(Electric Power Research Institute of Yunnan Power Grid Corporation,Kunming 650127,Yunnan,China;Yunnan Power Grid Co.,Ltd.,Chuxiong Lufeng Power Supply Bureau,Chuxiong 651200,Yunnan,China;North China Electric Power University(Baoding),Baoding 071000,Hebei,China;Yunnan Power Grid Co.,Ltd.,Dali Power Supply Bureau,Dali 671000,Yunnan,China;Yunnan Power Grid Co.,Ltd.,Qujing Power Supply Bureau,Qujing 655500,Yunnan,China;Yunnan Power Grid Co.,Ltd.,Pu'er Power Supply Bureau,Pu'er 665000,Yunnan,China)

机构地区：[1]云南电网有限责任公司电力科学研究院,云南昆明650127 [2]云南电网有限责任公司楚雄禄丰供电局,云南楚雄651200 [3]华北电力大学(保定),河北保定071000 [4]云南电网有限责任公司大理供电局,云南大理671000 [5]云南电网有限责任公司曲靖供电局,云南曲靖655500 [6]云南电网有限责任公司普洱供电局,云南普洱665000

出　　处：《云南电力技术》2024年第6期75-80,共6页Yunnan Electric Power

基　　金：云南电科院、曲靖供电局、大理供电局、普洱供电局2023年10kV低压台区预测和无功补偿容量研究及移动式新型10kV调压成套装置试制(技术服务部分).合同编号:0562002023030301XT00016。

摘　　要：提出了一种基于CNN-LSTM-GNN混合模型的10 kV台区低电压中长期预测方法。首先,对云南地区的电力负荷数据进行分析,考虑了季节性变化、温度变化、社会安全事件及线路配置等因素对负荷预测的影响。随后,构建了一种结合卷积神经网络(CNN)、长短期记忆网络(LSTM)和图神经网络(GNN)的深度学习模型,用于提高电力负荷及低电压预测的准确性。通过实验分析,模型在电力负荷预测及台区低电压治理决策支持方面显示出了良好的性能。特别是考虑到各类影响因素后,混合模型相较于传统方法在预测准确度与效率方面都有明显提高,其预测结果符合实际工程需求,说明了该方法的实用价值。本研究最后通过与其他技术的比较分析,进一步验证了所提方法的优越性。Based on the analysis of factors such as seasonal and temperature changes,social safety events,and line configuration affecting the 10kV sub-district low voltage in Yunnan,China,this study proposes a mid-to-long term prediction method employing a CNN-LSTM-GNN hybrid model.Recognizing the impact of seasonal variations,tourism fluctuations due to pandemics,and the topology of power distribution networks,the research integrates Convolutional Neural Networks(CNN)for extracting temporal and seasonal load features,Long Short-Term Memory(LSTM)networks for learning and forgetting irrelevant features over time,and Graph Neural Networks(GNN)for understanding the complex relational data among different network nodes.Through comprehensive data preparation and model training using historical load,temperature,transmission distance,and holiday information as well as social event impacts,the proposed model demonstrates remarkable voltage prediction accuracy with nearly 100%within a 1V error margin in the simulation,significantly outperforming traditional methods in terms of feature requirement,computational complexity,and accuracy.This approach not only offers a robust prediction model for managing low voltage issues across sub-districts but also proves efficient in extrapolating from power data to assess voltage situations across the entire network topology without needing extensive operational parameters.

关 键 词：电力负荷预测 低电压预测 深度学习 CNN-LSTM-GNN混合模型 

分 类 号：TM74[电气工程—电力系统及自动化]