基于混合卷积神经网络的多特征负荷预测方法研究  

Research on the Multi Feature Load Forecasting Method Based on Hybrid Convolutional Neural Network

作　　者：邹晴 李乐 柳楠 李超然 曹竞元 于金骁 朱霄珣[2] 于淼 ZOU Qing;LI Le;LIU Nan;LI Chaoran;CAO Jingyuan;YU Jinxiao;ZHU Xiaoxun;YU Miao(State Grid Beijing Daxing Electric Power Supply Company,Beijing 102600,China;North China Electric Power University,Baoding 071003,Hebei,China)

机构地区：[1]国网北京大兴供电公司,北京102600 [2]华北电力大学,河北保定071003

出　　处：《电网与清洁能源》2024年第9期54-62,共9页Power System and Clean Energy

基　　金：国家自然科学基金面上项目(52076081)。

摘　　要：针对负荷预测任务中准确性、稳定性和环境因素适应性的挑战,提出了一种基于混合卷积神经网络的电力负荷短期预测方法。提出了基于一维卷积神经网络(1D convolutional neural network,1D-CNN)的多尺度特征融合方法,通过融合不同尺度的特征来捕捉负荷变化的趋势,提高了对负荷突变和复杂模式的识别能力;针对多种环境特征因素对电负荷影响的问题,设计了基于2D-CNN的多特征因素学习方法,提高了模型对环境因素与负荷间复杂关系的建模能力;构建了混合网络模型,通过对1D-CNN和2D-CNN的特征信息进行深度特征融合和信息传播,实现了有效关联时空特征的综合性负荷预测方法。开展了具体算例分析研究,通过分析参数优化和融合学习对模型精度和效率的影响,并与经典模型进行对比,结果显示所提模型的均方根误差(root mean squared error,RMSE)为36.3,平均绝对误差(mean absolute error,MAE)为5.34,平均绝对百分比误差(mean absolute percentage error,MAPE)为1.02%,有效提高了负荷预测的准确性和鲁棒性。This paper proposes a short-term power load forecasting method based on hybrid convolutional neural networks to address the challenges of accuracy,stability,and adaptability to environmental factors in load forecasting tasks.First of all,A multi-scale feature fusion method based on 1D-CNN(1D convolutional neural network,1D-CNN)is proposed,which captures the trend of load changes by fusing features of different scales,improving the recognition ability of load mutations and complex patterns;A multi feature factor learning method based on 2D-CNN is designed to address the impact of various environmental characteristic factors on electricity loads,which improves the modeling ability of the model for complex relationships between environmental factors and loads.Second,a hybrid network model is constructed to achieve a comprehensive load forecasting method that effectively associates spatiotemporal features through deep feature fusion and information propagation of 1D-CNN and 2D-CNN feature information.Specific case studies are conducted to analyze the impact of parameter optimization and fusion learning on model accuracy and efficiency,and compared with classical models.The results show that the root mean squared error(RMSE)value of the model is 36.3,while the mean absolute error value is 5.34,and the mean absolute percentage error(MAPE)value is 1.02%,effectively improving the accuracy and robustness of the load forecasting.

关 键 词：负荷预测 混合卷积神经网络 多尺度特征融合 多特征因素 融合学习 

分 类 号：TM714[电气工程—电力系统及自动化]