基于图形处理器的广义最小残差迭代法在电力系统暂态仿真中的应用  

Application of Graph Processing Unit-Based Generalized Minimal Residual Iteration in Power System Transient Simulation

作　　者：唐聪[1] 严正[1] 周挺辉[1] 

机构地区：[1]电力传输与功率变换控制教育部重点实验室(上海交通大学),上海市闵行区200240

出　　处：《电网技术》2013年第5期1365-1371,共7页Power System Technology

基　　金：高等学校博士学科点专项科研基金资助项目(20120073110020);国家电网公司科技项目(X717201200041)~~

摘　　要：文中对电力系统暂态仿真算法及并行化设计进行了研究,针对图形处理器(graphic processing unit,GPU)的特性,应用广义最小残差法(generalized minimal residual,GMRES)提出一种基于GPU的电力系统暂态仿真并行算法。该算法采用预处理算法对暂态仿真计算过程中的系数矩阵进行预处理,降低条件数以提升收敛速度,经预处理后的线性方程组通过GMRES算法在GPU上并行求解,针对暂态仿真计算中线性方程组稀疏性的特点,算法应用稀疏存储技术以节省计算量和内存占用空间。测试表明,所提出的GPU并行算法与PSAT软件计算结果近似;相对CPU串行程序,当算例规模足够大时,GPU并行算法的加速效果明显,实测最高加速比为3.3。The algorithm for transient simulation of power system and its parallelization design are researched. Based on the features of graphic processing unit （GPU）and applying generalized minimal residual （GMRES）, a GPU-based parallel algorithm for power system transient simulation is proposed. In the proposed algorithm, the pretreatment algorithm is utilized to pretreat the coefficient matrix in transient simulation process to reduce the numbers of conditions to speed up the convergence; by means of GMRES algorithm the pretreated linear equations are solved by GPU. According to the sparsity of linear equations during transient simulation, the sparse storage technique is adopted to save the memory footprint and decrease the computing capacity. Test results show that the computing results by the proposed GPU parallel algorithm are similar to those by PSAT software; relative to CPU serial algorithm, when the scale of the calculation example is big enough, the accelerating effect of GPU parallel algorithm is remarkable, and the measured highest acceleration ratio is 3.3.

关 键 词：暂态仿真 预处理 广义最小残差法 图形处理器 并行计算 统一计算设备架构 稀疏技术 

分 类 号：TM744[电气工程—电力系统及自动化]