电量预测中的多分辨时序神经网络模型研究  

Multi-resolution time sequence neural net model for electricity consumption forecasting

作　　者：石少青 郑楷洪 周尚礼 杨劲锋 张英楠 陈敏娜 张健[3] SHI Shaoqing;ZHENG Kaihong;ZHOU Shangli;YANG Jinfeng;ZHANG Yingnan;CHEN Minna;Zhang Jian(China Southern Power Grid Co. Ltd., Guangzhou, Guangdong 510663, China;China Southern Power Grid Digital Grid Research Institute Co. Ltd., Guangzhou, Guangdong 510663, China;China University of Mining and Technology, Xuzhou, Jiangsu 221116, China)

机构地区：[1]中国南方电网有限责任公司,广东广州510663 [2]南方电网数字电网研究院有限公司,广东广州510663 [3]中国矿业大学,江苏徐州221116

出　　处：《山东科技大学学报（自然科学版）》2022年第2期117-126,共10页Journal of Shandong University of Science and Technology(Natural Science)

基　　金：南方电网科技项目(670000KK52200081)。

摘　　要：精确的电量预测可以为电量生产和供电调度提供可靠的指导,不同行业的用电量数据分布特性不同,用电量的周期性也因行业周期的不同具有不同时间尺度。传统的用电量预测模型难以在拟合用电量数据周期性的同时捕捉数据的波动性,为此本研究提出了多分辨时序神经网络(MTNN)模型,根据用电量数据特性将用电量数据视为近似的周期性数据分量和波动分量的组合,从多时频分辨率的角度对用电量数据进行多尺度分解,在不同时频尺度下预测用电量数据分量,然后通过组合预测分量实现用电量预测任务。具体而言,为捕捉数据周期性,首先使用集成经验模态分解方法(EEMD)对用电量数据分解,得到不同时间尺度的模态分量(IMF),使用LSTM预测高阶IMF分量;对于低阶IMF分量,设计了多阶时频注意力神经网络计算低阶IMF分量的预测值,最后组合多个IMF分量的预测值得到电量预测结果。实验结果表明,与常用的电量预测模型相比,MTNN能更好地捕捉电量数据的变化规律,提高电量预测的精度。Accurate electricity consumption forecasting provides a reliable guidance for electricity production.However,the characteristics of electricity consumption data in different industries are different.But,rare electricity consumption forecasting models are specially designed to capture the fluctuation and the periodicity of electricity consumption data.To effectively predicting electricity consumption,this paper proposed a multi-resolution time sequence neural network(MTNN),regarding the electricity consumption data as combinations of approximate periodic components and fluctuation components.Specifically,to capture the periodicity of the data,the MTNN decomposed the power consumption data by using the ensemble empirical mode decomposition(EEMD)method to obtain intrinsic mode functions(IMF)in different time scales.Then,LSTM was used to predict the high-order IMF components.For the low-order IMF components,the MTNN designed a multi-order time-frequency attention neural network to calculate the predicted value of the low-order IMF components.The experimental results show that compared with the commonly-used electricity consumption forecasting models,the MTNN can better capture the characteristics of electricity data,thus improving the accuracy of electricity consumption forecasting.

关 键 词：电量预测 多分辨分析 时序神经网络 时频神经注意力 

分 类 号：TM714[电气工程—电力系统及自动化]