基于GRU-Dropout网络的避雷器阻性电流预测研究  

Study of Lightning Arrester Resistive Current Prediction Based on GRU-Dropout Network

作　　者：刘谱辉 周国平[1,2] 周逸鹏 LIU Puhui;ZHOU Guoping;ZHOU Yipeng(School of Information Science and Technology,Nanjing Forestry University,Nanjing 210018,China;School of Artificial Intelligence,Nanjing Forestry University,Nanjing 210018,China)

机构地区：[1]南京林业大学信息科学技术学院,南京210018 [2]南京林业大学人工智能学院,南京210018

出　　处：《自动化与仪表》2024年第8期6-9,19,共5页Automation & Instrumentation

基　　金：国家自然科学基金资助项目(32171788)。

摘　　要：MOA(氧化锌避雷器)阻性电流在设备过电压时具有引导、分流和保护的作用,通过对其进行准确预测能够有效保证电力系统的安全与稳定。针对部分单一网络过拟合、泛化能力不足等弊端提出了一种基于GRU(门控循环单元)结合Dropout(随机失活)的MOA阻性电流预测方法。结合变电站实际电流数据,并设置合适的超参数对网络进行训练和分析。实验结果表明,采用GRU-Dropout网络相较于LSTM提高了泛化能力,减少了过拟合;相较于RNN和BP神经网络提高了精度,其最优平均绝对误差和均方根误差为1.98%和2.73%。在实际应用中能够较为准确地预测MOA对地绝缘趋势,具有一定的应用价值。MOA(metal oxide arrester)resistive current has the role of guiding,shunting and protecting during equipment overvoltage,and its accurate prediction can effectively ensure the safety and stability of the power system.A MOA resistive current prediction method based on GRU(gated recurrent unit)combined with Dropout(stochastic deactivation)is proposed to address the drawbacks of overfitting and insufficient generalization ability of some single networks.The network is trained and analyzed by combining the actual current data of the substation and setting appropriate hyperparameters.The experimental results show that the use of GRU-Dropout network improves the generalization ability and reduces overfitting compared with LSTM.Improves the accuracy compared with RNN and BP neural networks,with its optimal average absolute error and root mean square error of 1.98%and 2.73%.It can predict the trend of MOA insulation to ground more accurately in practical applications,and has certain application value.

关 键 词：氧化锌避雷器 阻性电流 门控神经网络 随机失活 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置]