GPU加速潮流算法的多重并行度挖掘及规则化重构策略研究

The new generation of dispatching system needs to make analysis on the entire power grid, which greatly increases the computational complexity of power flow (PF). Some applications, such as static security analysis and probabilistic power flow, need tens of thousands of PF calculation. Therefore, it is necessary to employ high performance computing technologies, such as GPUs, to speed up the PF analysis. In order to reformulate the PF calculation into a large-scale regular computing problem that GPUs excel in, the applicant focuses on how to exploit multiple and regular parallelism for GPU-accelerated PF algorithm. Firstly, we extract various kinds of parallelisms from the stratified elimination tree, sparse vector operations and the batch PF calculation, and establish the high-efficiency mapping strategies between aforementioned parallelisms and GPU threads. On this basis, by the means of unifying the sparse pattern and the joint optimization of memory and thread allocations, the batch PF computing problems, such as static security analysis, can be reconstructed into a regular dense vector algorithm with the characteristics of solving a large number of PFs concurrently, unified thread logic and continuous memory access. At the same time, we propose the parameterized performance evaluation and design methods of GPU-based algorithms, which guide the design, optimization and application of GPU-based PF algorithm. This study will propose a set of design methodologies and key algorithms for GPU-accelerated PF calculation, which maybe speed up PF analysis 1 to 2 orders of magnitude and provide a new idea of applying GPU in power system.

我国新一代调度系统要求进行全网统一分析，这大幅增加了潮流的计算复杂度，静态安全分析、概率潮流等应用更是需要计算大批量潮流，有必要采用GPU等高性能计算技术来加速潮流分析。为将潮流重构为GPU擅长的并行度高、线程束收敛、访存合并的大型规则化计算问题，本申请围绕多重并行度挖掘和规则化重构方法开展研究：首先从分层依赖树、稀疏向量运算、批量潮流中挖掘多重并行度，并建立各类并行度与GPU线程的高效映射策略；在此基础上，从稀疏格式统一设计、内存和线程分配的联合优化入手，将静态安全分析等大量潮流计算问题重构为一个同时计算批量潮流、逻辑统一、访存连续的规则化稠密向量算法；同时，提出GPU算法的参数化性能分析和设计方法，指导算法的设计、优化和应用全过程。项目研究将提出一套GPU加速潮流计算的设计理论和关键算法，将潮流的分析速度提升1~2个数量级，为GPU在电力系统中的应用提供思路。

随着新一代调控系统与调控云的开发，电网建模和数据采集向配网扩展，全网稳态分析应用的计算范围不断扩大，故障预想更加全面，需要在保证计算精度的前提下进一步提升网络分析计算方法的时效性。传统模型和算法简化的处理方法难以兼顾计算精度和速度，分区分块并行算法的并行度和 CPU 并行计算的能力也存在一定局限。而图形处理器（Graphics Processing Unit， GPU）凭借其强大的计算能力成为通用并行计算领域不可忽视的一股力量，在电力系统稳态安全分析的应用中有着巨大潜力。在项目执行期间，项目组提出了一套GPU加速潮流计算的设计理论和关键算法，将潮流的分析速度提升1~2个数量级，为GPU在电力系统的深度应用提供了思路。主要研究内容包括：1）挖掘分析了GPU潮流算法的多重并行度分解和映射策略，为单个潮流加速提供了解决方案；2）在单个潮流加速的基础上深入研究了批量潮流问题的规则化重构和优化设计，进一步提升了稳态分析应用的计算效率；3）开展了GPU批量潮流算法在静态安全分析、概率潮流中的深度应用，为GPU在新一代调度控制系统中的实际应用提供依据。