Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)


School of Computing

Committee Chair/Advisor

Shuangshuang Jin

Committee Member

Amy Apon

Committee Member

Rong Ge

Committee Member

Ramtin Hadidi


This dissertation presents an HPC-enabled fast and configurable dynamic simulation, analysis, and learning framework for complex power system adaptation and control. Dynamic simulation for a large transmission system comprising thousands of buses and branches implies the latency of complicated numerical computations. However, faster-than-real-time execution is often required to provide timely support for power system planning and operation. The traditional approaches for speeding up the simulation demand extensive computing facilities such as CPU-based multi-core supercomputers, resulting in heavily resource-dependent solutions. In this work, by coupling the Message Passing Interface (MPI) protocol with an advanced heterogeneous programming environment, further acceleration can be achieved up to 14.0x faster than the sequential version using limited computing hardware.

Dynamic contingency analysis helps system planning engineers and operators assess the impact and likelihood of potential cascading events. In the current practice, contingency tasks are performed in a serial manner. In this work, a configurable hierarchical architecture is developed for parallel massive dynamic contingency analysis to handle ‘what if’ questions simultaneously on multi-GPUs. The contingency cases are distributed to different GPUs for running specific time-domain simulations within CUDA kernels. Compared to the CPU-based distributed solution, our test shows an up to 2.8x speedup using one GPU and a 4.2x speedup using two GPUs, respectively. The promising results provide concrete evidence of HPC's significant role in improving computational performance without compromising computational accuracy.

Deep reinforcement learning in Artificial Intelligence has been introduced to predict and control the behavior of load-shedding in complex power systems. To investigate the parameter space of the control policy effectively and be suited to understand multiple fault scenarios in the power system, the algorithms need to perform many power system dynamic simulations by considering a large number of different perturbed policies at each training iteration. This work discusses the distributed Augmented Random Search algorithm powered by a robust simulation model, which implements the high-efficient reduced admittance matrix method on the CPU. With advanced mathematical optimization and minimized data transfer, the training time with nine scenarios and 300 iterations can achieve a 1.4x speedup compared to the original approach.



To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.