基于卷积神经网络的负荷预测  

Load forecasting based on convolutional neural networks

作　　者：李德璐 赵金脉 李大华[1] 田禾[1] LI De-lu;ZHAO Jin-mai;LI Da-hua;TIAN He(School of Electrical Engineering and Automation,Tianjin University of Technology,Tianjin 300384,China;Design and Development Center of CNOOC Energy Development Equipment Technology Company Limited,Tianjin 300384,China)

机构地区：[1]天津理工大学电气工程与自动化学院,天津300384 [2]中海油能源发展装备技术有限公司设计研发中心,天津300384

出　　处：《能源工程》2022年第6期75-79,共5页Energy Engineering

摘　　要：电力负荷预测对于保证大扰动下的系统稳定、优化智能电网中的能源分配具有重要意义。传统的预测模型主要基于时间序列分析,由于其不可忽略的预测误差,已经不能完全满足电力系统的实际需要。为提高预测精度,将时间序列数据分析转变为图像处理,并利用计算机图像领域广泛使用的深度学习方法进行电力负荷预测。卷积神经网络(convolution neural network,CNN)作为图像处理的有力工具,尽管已有学者将其用于时间序列数据处理,但仍是将数据作为序列矩阵处理,并未体现出CNN处理图像矩阵的优势。因此提出基于序列到图像转换的CNN(sequence to image convolutional neural network,STI-CNN),将负荷序列转换为负荷图像,使CNN可以更有效地提取相邻信息特征,充分考虑到各种外部影响因素,使用双分支深度网络模型对输入数据进行精确聚类,通过STI-CNN方法进行负荷预测。负荷预测实验结果表明,所提STI-CNN方法在不同的预测指标方面都有卓越的表现,所用预测时间更短,具有更高准确度。Electricity load forecasting is great importance for ensuring system stability under large disturbances and optimizing energy distribution in smart grids.Traditional forecasting models,mainly based on time series analysis,no longer fully meet the practical needs of power systems due to their non-negligible forecasting errors.To improve prediction accuracy,time series data analysis is transformed into image processing and deep learning methods,which are widely used in the field of computer graphics,are utilized for power load prediction.Convolution neural network(CNN)is used as a powerful tool for image processing.Although CNN has been used for time series data processing,it still treats the data as a sequence matrix,which does not reflect the advantages of CNN for image matrix processing.Therefore,CNN based on Sequence to image convolution neural network(STI-CNN for short)was proposed in this paper,which converts load sequences into load images so that the CNN can extract adjacent information features more effectively,fully taking into account various external influencing factors,and using a two-branch deep network model the input data is accurately clustered and load prediction is performed by the STI-CNN method.The experimental results of load prediction show that the proposed STI-CNN method has excellent performance in different prediction metrics,and the prediction time used is shorter and has higher accuracy.

关 键 词：短期负荷预测 卷积神经网络 负荷序列 负荷图像 

分 类 号：TM715[电气工程—电力系统及自动化]