基于多尺度时间特征LSTM的短期负荷预测  

Short-term Load Prediction of LSTM Based on Multi-scale Time Characteristics

作　　者：杨梅[1] 李忠 吴昊[3] 代妮娜[1] YANG Mei;LI Zhong;WU Hao;DAI Ni-na(Key Laboratory of Information and Signal Processing,Chongqing Three Gorges University,Chongqing 404130,China;State Grid Ziyang Power Supply Company,Ziyang 641300,China;State Grid Changzhou Power Supply Company,Changzhou 213000,China)

机构地区：[1]重庆三峡学院信息与信号处理重点实验室,重庆万州404130 [2]国网资阳供电公司,四川资阳641300 [3]国网常州供电公司,江苏常州213000

出　　处：《控制工程》2022年第9期1722-1728,共7页Control Engineering of China

基　　金：重庆市教委科学技术研究项目(KJQN201801213)。

摘　　要：为了提高样本数据单一情况下的负荷预测精度,提出了基于多尺度时间特征的长短时记忆网络LSTM模型。首先,采用小波分解将历史数据分解为稳定分量、趋势负荷、以及峰-谷周期和持续时间等周期序列,突出不同时间尺度特征;其次,利用LSTM网络实现时间序列特性的进一步提取和数据拟合;最后,模型直接输出多个时刻的预测值。实验表明,相比较于自组织映射、高斯过程回归、标准LSTM模型,所提基于多尺度时间特征的长短时记忆网络模型具有更高的预测精度,同时具有一定的抗噪性能。In order to improve the accuracy of load prediction with single sample data, a long short-serm memory(LSTM) model based on multi-scale time characteristics is proposed. Firstly, wavelet decomposition is used to decompose the historical data into periodic series such as stable component, trend load, size of peak-valley period and duration time, highlighting the characteristics of different time scales. Secondly, the LSTM network is used to further extract the characteristics of time series and realize data fitting. Finally, the model directly outputs the predicted values of multiple moments. Experiments show that compared with self-organizing mapping, Gaussian process regression and standard LSTM model, the proposed LSTM model based on multi-scale time characteristics has higher prediction accuracy and certain anti-noise performance.

关 键 词：时间序列 负荷预测 长短时记忆网络 数据分解 

分 类 号：TM71[电气工程—电力系统及自动化]