基于GPU的分块约化算法在小干扰稳定分析中的应用

为了提高电力系统小干扰稳定全部特征值分析的计算速度,研究了QR算法中上Hessenberg约化算法的并行化方法。以分块的方式将约化算法中的浮点运算整合为高阶的基础线性代数子程序(BLAS)运算,实现了分块约化算法在中央处理器(CPU)/图形处理器(GPU)混合架构下的并行,并应用到大规模电力系统的小干扰稳定全部特征值分析中。仿真结果表明,相比于多核CPU并行,基于GPU的分块上Hessenberg约化算法取得了高达5倍的加速效果。包含所提方法的全部特征值分析的整体计算速度获得了显著的提升,提高了QR算法对于大规模电力系统仿真分析的适用性。

Application of GPU-based Block Reduction Algorithm in Power System Small-signal Stability Analysis

To enhance the computational efficiency of complete eigenvalue analysis in power system small-signal stability analysis, the parallelization of upper Hessenberg reduction algorithm in the QR method is studied. A block reduction algorithm is utilized to integrate the floating-point operations into high-level basic linear algebraic subprograms (BLAS). The block reduction algorithm is parallelized on hybrid CPU/GPU (graphic processing unit) system and applied to the complete eigenvalue analysis of large-scale power system small-signal stability analysis. Simulation results show that, compared with multi-core CPU parallelization, the GPU-based block upper Hessenberg reduction algorithm is able to obtain a speed-up ratio up to 5 times the original. The overall computing speed of the complete eigenvalue analysis, including the method proposed, has achieved remarkable acceleration improvement. The applicability of the QR method to large-scale power system simulation analysis is increased. This work is supported by State Grid Corporation of China, Major Projects on Planning and Operation Control of Large Scale Grid (No. SGCC-MPLG018-2012) and Specialized Research Fund for the Doctoral Program of Higher Education (SRFDP) of China (No. 20120073110020).