基于Xception卷积与权重剪枝的轻量级短期负荷预测方法  

Lightweight Short-term Load Forecasting Method Based on Xception Convolution and Weight Pruning

作　　者：吴涵 范元亮 林建利 李泽文 李凌斐 黄兴华 WU Han;FAN Yuanliang;LIN Jianli;LI Zewen;LI Lingfei;HUANG Xinghua(Electric Power Research Institute of Fujian Electric Power Co.,Ltd.,Fuzhou 350000,China)

机构地区：[1]国网福建省电力有限公司电力科学研究院,福建福州350000

出　　处：《电工技术》2024年第16期86-90,94,共6页Electric Engineering

基　　金：国网福建省电力有限公司科技项目(编号52130422003N)。

摘　　要：精准而快速的负荷预测能够促进电力系统的经济稳定运行。为使负荷预测模型能够应用到资源受限的边端台区并保证良好的预测精度,提出了一种基于Xception卷积神经网络与权重剪枝的轻量级负荷预测模型。首先,选取历史数据集中的日类型、历史负荷、温度及湿度数据作为模型输入并将其转化为特征矩阵;然后,根据输入特征矩阵和模型输出向量的特点,基于Xception卷积与注意力机制构建轻量级台区负荷预测模型;最后,基于权重剪枝的模型训练方法生成一个轻量化的负荷预测模型。通过对比实验发现,所提出的轻量化负荷预测模型的参数量、计算量及存储空间相比大型模型大幅度降低,而其精度与大型模型基本相当,在边端应用中更具有优势。Accurate and fast load forecasting can promote the economic and stable operation of the power system.In order to make load forecasting model applicable to the resource-constrained side-terminal stations and to ensure good forecasting accuracy,a lightweight load forecasting model based on Xception convolutional neural network with weight pruning is proposed.First the daily type,historical load,temperature and humidity data from the historical data set are selected as model inputs and transformed into feature matrices.Then a lightweight load forecasting model is constructed based on the Xception convolution and the attention mechanism according to the characteristics of the input feature matrices and the output vectors of the model.Finally a lightweight load forecasting model is established by weight pruning-based model training.Through comparative experiments,it is found that the number of parameters,computational volume and storage space of the lightweight load forecasting model in this paper are greatly reduced,while its accuracy is basically comparable to that of the large-scale model,which is more advantageous in side-end applications.

关 键 词：负荷预测 Xception卷积 剪枝 轻量级模型 

分 类 号：TM715[电气工程—电力系统及自动化]