Research on Dimensionless Least Square Finite Element Method GPU Implementation and Its Application in Transformer Winding Flow Field Simulation
靳立鹏 1刘刚 1任增强 1李浩 1武卫革2
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 河北省输变电设备安全防御重点实验室(华北电力大学),河北保定 071003
2. 保定天威保变电气股份有限公司河北省输变电装备电磁与结构性能重点实验室,河北保定 071056
折叠
摘要
为了提高变压器流体场的计算效率,利用统一计算设备架构(compute unified device architecture,CU-DA)实现流场的并行计算.首先基于C语言实现无量纲最小二乘有限元法的串行计算方法.然后在Visual Studio 2019+CUDA 10.2的环境下实现并行运算,对于串行程序中最耗时的线性方程组求解部分调用了CU-DA自带的函数库进行优化;在大规模模型的数值计算中使用了十字链表格式存储整体刚度阵形成时的非零元素,以解决满阵存储时的内存不足问题.为验证所提方法的有效性,针对方腔模型,分析了不同剖分网格数量下的加速比,分析结果表明,数据规模越大并行效果越好,验证了无量纲最小二乘有限元并行程序的准确性和高效性.最后将该方法应用到大型变压器绕组的流体场分析中,取得了18.3倍的加速效果,为产品级变压器的流体场仿真奠定了基础.
Abstract
In order to improve the computational efficiency of transformer fluid field,this paper studies the parallel computation of flow field using compute unified device architecture(CUDA).Firstly,we implemented the serial com-putation method of the dimensionless least squares finite element method based on C language.Then,the parallel com-putation was implemented in Visual Studio 2019+CUDA 10.2 environment,and for solving the most time-consuming linear equations in the serial program,the CUDA built-in function library was called to optimize.In the numerical cal-culation of large-scale model,the non-zero elements of the global stiffness matrix are stored in a tabular form to solve the memory footprint problem.In order to verify the effectiveness of the proposed method,we analyzed the acceleration ratio under different number of meshes for the square cavity model.The analysis results show that the larger the data scale,the better the parallel effect,which verifies the accuracy and efficiency of the dimensionless least squares finite element parallel program.Finally,we applied the method to the fluid field analysis of large transformer windings,and the acceleration effect is 18.3 times,which lays a foundation for the fluid field simulation of product-grade transformer.
关键词
变压器/绕组/无量纲最小二乘有限元/流场/GPU/加速比
Key words
transformer/winding/dimensionless least squares finite element/fluid field/GPU/speed-up ratio
维普
万方数据