System-level electromagnetic transient (EMT) simulation of large-scale power converters with high-order nonlinear semiconductor switch models remains a challenge albeit it is essential for design preview. In this work, a multi-layer hierarchical modeling methodology is proposed for high-performance computing of the modular multilevel converter involving device-level IGBT/diode models. The computational burden induced by converter scale and model complexity is dramatically alleviated following the proposal of topological reconfiguration and network equivalence, which create a substantial number of identical circuit units that facilitate massively parallel processing on the graphics processing unit (GPU), using the kernel-based single-instruction multi-threading computing architecture. As the DC system brings significant inhomogeneity which dilutes parallelism, heterogeneous computing is investigated and the computational tasks are properly assigned to CPU and GPU to fully exploit their respective features. The separation of nonlinear device-level models from the rest of the system enables multi-rate implementation for further efficiency enhancement since the two parts allow distinct time-steps. A remarkable acceleration of over 50 times is achieved by the hybrid CPU/GPU platform over conventional CPU simulation, and the validity of the proposed modeling and computing method is confirmed by commercial EMT tools ANSYS/Simplorer and PSCAD/EMTDC.

中文翻译：

---

高压直流输电系统并行和异构瞬态仿真的分层设备级模块化多级转换器建模

具有高阶非线性半导体开关模型的大型功率转换器的系统级电磁瞬态（EMT）仿真仍然是一个挑战，尽管这对于设计预览至关重要。在这项工作中，提出了一种多层层次建模方法，用于对包含器件级IGBT /二极管模型的模块化多级转换器进行高性能计算。提出了拓扑重新配置和网络等效性的提议后，由转换器规模和模型复杂性引起的计算负担得到了极大的缓解，该提议使用内核来创建大量相同的电路单元，从而促进了图形处理单元（GPU）上的大规模并行处理。基于单指令多线程的计算架构。由于DC系统带来显着的不均匀性，从而降低了并行性，因此对异构计算进行了研究，并将计算任务正确分配给CPU和GPU，以充分利用它们各自的功能。非线性设备级模型与系统其余部分的分离实现了多速率实施，从而进一步提高了效率，因为这两部分允许不同的时间步长。与传统的CPU仿真相比，混合CPU / GPU平台可实现超过50倍的显着加速，而商用EMT工具ANSYS / Simplorer和PSCAD / EMTDC证实了所提出的建模和计算方法的有效性。非线性设备级模型与系统其余部分的分离实现了多速率实施，从而进一步提高了效率，因为这两部分允许不同的时间步长。与传统的CPU仿真相比，混合CPU / GPU平台可实现超过50倍的显着加速，而商用EMT工具ANSYS / Simplorer和PSCAD / EMTDC证实了所提出的建模和计算方法的有效性。非线性设备级模型与系统其余部分的分离实现了多速率实施，从而进一步提高了效率，因为这两部分允许不同的时间步长。与传统的CPU仿真相比，混合CPU / GPU平台可实现超过50倍的显着加速，而商用EMT工具ANSYS / Simplorer和PSCAD / EMTDC证实了所提出的建模和计算方法的有效性。