基于改进卷积神经网络的新能源并网短路电流预测技术  

Short circuit current prediction technology for new energy connected to the power grid based on improved convolutional neural network

作　　者：于琳琳 蒋小亮 贾鹏 孟高军[2] 丁咚 Yu Linlin;Jiang Xiaoliang;Jia Peng;Meng Gaojun;Ding Dong(State Grid Henan Electric Power Company Economic and Technological Research Institute,Zhengzhou 450052,China;School of Electrical Engineering,Nanjing Institute of Technology,Nanjing 211167,China)

机构地区：[1]国网河南省电力公司经济技术研究院,河南郑州450052 [2]南京工程学院电力工程学院,江苏南京211167

出　　处：《可再生能源》2025年第3期408-415,共8页Renewable Energy Resources

基　　金：江苏省重点研发计划(BE2021094)。

摘　　要：随着新能源的大规模并网,大电网短路电流特征更加复杂、难以预测。基于此,文章提出了一种基于改进卷积神经网络的新能源并网短路电流预测技术。首先,分析短路电流特点,对短路电流进行变分模态分解,得到本征模态函数;其次,对卷积神经网络进行改进,利用多尺度特征提取将电流故障数据特征最大化,引入注意力机制提取重要信息,卷积过程中使用跳跃连接的方式防止前向传递时信息丢失,有利于提高预测的准确性,构建基于改进卷积神经网络的短路电流预测模型;最后,经过PSCAD/EMTDC电网模型进行验证。结果表明,所提方法对短路电流峰值预测有着较高的精度,与常见的极限学习机、支持向量机相比,平均相对误差分别降低了0.61%,1.09%,验证了文章所提方法的有效性。With the large-scale integration of distributed power sources, the short-circuit current characteristics of large power grids become more complex and difficult to predict. Based on this,this article proposes a new energy grid short-circuit current prediction technology based on improved convolutional neural networks. Firstly, analyze the characteristics of short-circuit current, perform variational mode decomposition on short-circuit current, and obtain the intrinsic mode function;Secondly, the convolutional neural network is improved by utilizing multi-scale feature extraction to maximize the features of current fault data, introducing attention mechanisms to extract important information, and using skip connections during the convolutional process to prevent information loss during forward transmission, which is beneficial for improving the accuracy of prediction. A short-circuit current prediction model based on the improved convolutional neural network is constructed;Finally, the PSCAD/EMTDC power grid model was validated, and the experimental results showed that the proposed method has high accuracy in predicting the peak short-circuit current. Compared with common limit learning machines and support vector machines, the average relative error decreased by 0.61% and 1.09%, respectively.This verified the effectiveness of the proposed method and laid the foundation for limiting short-circuit current in large power grids.

关 键 词：新能源 改进卷积神经网络 短路电流预测 变分模态分解 注意力机制 

分 类 号：TK51[动力工程及工程热物理—热能工程] TK81[电气工程—电力系统及自动化] TM73