基于道路树分层的大电网潮流并行算法及其GPU优化实现

针对大规模电网分析及能量管理系统对快速潮流计算的需求,提出了一种适于图形处理器(GPU)的基于道路树分层的稀疏矩阵直接分解算法,并结合该算法在GPU上实现了基于牛顿-拉夫逊法的潮流计算.为提高基于GPU的计算效率,首先在GPU上实现了潮流方程式右端项生成、雅可比矩阵生成、LU分解以及前推回代求解,减少了CPU和GPU之间的数据传输时间.其次,针对GPU中寄存器-缓存-显存多级存储架构,改进数据存储方式,减少了读取延迟.进一步,考虑GPU线程组织特点,优化任务分配,增加了计算并行度.最后,对比基于CPU的电力系统分析综合程序(PSASP)潮流计算模块,进行了数值仿真测试.结果表明,随着节点数的增加,所提出的程序计算优势越来越显著,算例规模达到43 602个节点时可获得5.172倍的加速比,验证了算法的有效性和实用性.

A Parallel Power Flow Algorithm for Large-scale Grid Based on Stratified Path Trees and Its Implementation on GPU

For the demand on fast power flow calculation for large-scale power grid analysis and energy management systems, a sparse matrix factorization algorithm for graphic processing unit (GPU) based on stratified path trees is proposed. By referring to the algorithm, the Newton-Raphson method for solving the power-flow problem is implemented on GPU. To improve the efficiency, the calculation of the right-hand side of equation and the Jacobian matrix, LU decomposition and forward-back substitution are implemented on GPU, which reduces the data transfer time between GPU and central processing unit (CPU). Then, considering the multi-level memory architecture with registers, cache and global memory, a special data structure is designed to decrease the latency of data access. Moreover, considering the feature of threads arrangement on GPU, the task allocation is optimized, while the efficiency of parallelism is improved. Finally, compared with the power flow calculation module of power system analysis software package (PSASP) ran on CPU, numerical simulation test is conducted. The results show that the more the nodes, the more advantages GPU will get. In the test on a system with 43 602 nodes, the program is able to provide a 5.172 times speed-up ratio, which proves the effectiveness and practicality of the algorithm proposed.