基于SCI-MN-SE模型的电力负荷预测方法  

Electricity load forecasting method based on SCI⁃MN⁃SE model

作　　者：李林骏 魏延[1] 李雪 谢渝 LI Linjun;WEI Yan;LI Xue;XIE Yu(College of Computer and Information Science,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学计算机与信息科学学院,重庆401331

出　　处：《电子设计工程》2025年第2期22-27,共6页Electronic Design Engineering

基　　金：重庆市技术创新与应用发展专项重点项目(cstc2019jscx-mbdxX0061)。

摘　　要：针对现有电力负荷预测方法对时间序列信息的提取效率较低、预测精度不高等问题,提出了一种基于样本卷积与交互网络(SCINet)和多层感知机(MLP)的电力负荷预测模型(SCI-MN-SE)。使用SCINet在不同时间分辨率下对负荷数据进行特征提取,从而更好地捕捉时间特征中的短期和长期信息关系;通过多层感知机叠加组成的瓶颈模块,沿着时间和特征维度进行混合操作来提高对时间序列信息的提取效率;结合通道注意力提高有效信息的权重,提升预测精度。在澳大利亚某地区真实电力负荷数据集上的实验结果表明,与TCN、Informer等方法相比,平均绝对百分比误差分别降低了26.30%、24.27%,证明SCI-MN-SE模型能有效地提取时间序列信息,预测效果较好。Aiming at the problems of low extraction efficiency of time series information and low accuracy of prediction results in existing power load forecasting methods,a power load forecasting model(SCI-MN-SE)based on Sample Convolution and Interaction Network(SCINet)and Multi⁃Layer Perceptron(MLP)is proposed.Feature extraction of load data at different temporal resolutions is performed using SCINet to better capture the relationship between short⁃term and long⁃term information in temporal features;The extraction efficiency of time series information is improved by a bottleneck module consisting of a superposition of multilayer perceptron machines,which performs a hybrid operation along the temporal and feature dimensions.The weighting of the effective information is improved by combining channel attention to enhance the prediction accuracy.The experimental results on the real power load dataset of a region in Australia show that the average absolute percentage error is reduced by 26.30%and 24.27%compared with TCN,Informer and other methods,respectively,which proves that the SCI-MN-SE model can extract the time series information efficiently and the prediction effect is better.

关 键 词：电力负荷预测 多层感知机 样本卷积与交互 通道注意力 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TN98[自动化与计算机技术—控制科学与工程]