摘要
在电气设备电磁场数值分析过程中,程序代码中含有大量的循环计算,其计算时间随有限元分析网格中单元和节点数量的增加而迅速增加。为提高循环计算时硬件资源的利用率、减少计算时间,根据循环计算相关特征量,研究能够预测不同循环中最佳循环展开因子的机器学习算法。首先建立不同的机器学习回归算法模型预测不同特征变量的循环计算时间,对模型的参数进行优化以提高预测准确率,通过比较计算时间找出最佳展开因子。然后建立循环特征变量与展开因子之间机器学习分类算法模型,预测得到最优循环展开因子。最后将算法应用于预测有限元分析代码中循环计算的最佳展开因子,将结果在1台单相变压器磁场有限元分析代码中进行循环展开,计算得到展开前后时间的加速程度。循环展开技术可以应用到有限元串行和并行计算中,并且在不同的程序中都有很好的适应性。
In the process of numerical analysis of electromagnetic fields of electrical equipment,the program code contains a large number of loop calculations.The calculation time increases rapidly with the increase of the number of elements and nodes in the finite element analysis grid.In order to improve the utilization of hardware resources and reduce the computing time during loop computation,according to the loop,the relevant characteristic quantity is calculated.It researches a machine learning algorithm that can predict the optimal loop expansion factor in different loops.Firstly,different machine learning regression algorithm models are developed to predict the loop computation time for different feature variables,the parameters of the models are optimised to improve the prediction accuracy,and the best unrolling factor is found by comparing the computation time.Then a machine learning classification algorithm model between the loop feature variables and the unrolling factor is developed to predict the optimal loop unrolling factor.Finally,the algorithm is applied to predict the best unrolling factor for the loop calculation in the finite element analysis(FEA)code,and the results are applied to the loop unrolling in a single-phase transformer magnetic field FEA code to calculate the acceleration in time before and after unrolling.The loop unrolling technique can be applied to both serial and parallel finite element calculations and well adapted to different programs.
作者
邢阔鹏
阎秀恪
任自艳
XING Kuopeng;YAN Xiuke;REN Ziyan(College of Electrical Engineering,Shenyang University of Technology,Shenyang,Liaoning 110870,China)
出处
《东北电力技术》
2024年第2期5-10,共6页
Northeast Electric Power Technology
基金
辽宁省研究生教育教学改革研究项目(LNYJG2022060)。
关键词
循环展开
电磁计算
机器学习
并行计算
loop unrolling
electromagnetic calculation
machine learning
parallel calculation
