Dynamic Modeling and Real-time Simulation of Power Electronics-dominated Power Grids Using Hybrid DDM and EDDM Techniques

Conducting transient stability simulations on complex and interconnected electrical power systems necessitates solving numerous differential algebraic equations. As modern grids grow in size and complexity, traditional sequential simulation methods have become increasingly computationally challenging and time-consuming. This research investigates parallelism techniques from both an innovative algorithmic perspective and at the hardware level, utilizing multiple graphics processing units (GPUs) to simulate the dynamic behavior of large-scale power systems. A novel hybrid algorithm is introduced to enhance dynamic simulations of grids with significant renewable energy integration. The algorithm employs a Schwarz-based approach in its initial decomposition phase to isolate synchronous machines and renewable sources, improving the management of these components. Each subsystem is further divided into subdomains, with separate computations performed using the Schur-complement technique. Simulations on test systems with up to 25,000 buses, 8,000 synchronous generators, and 256 PV farms demonstrate a notable 7.8x acceleration in speed when executed on an NVIDIA GeForce RTX 2070 SUPER GPU, thanks to a GPU-oriented preprocessing and vectorization parallelization method tailored to enhance the performance of the dynamic domain decomposition algorithm. This thesis also explores integrating an extended domain decomposition method (DDM) with machine learning, leading to significant advancements in Scientific Machine Learning (SciML). It introduces Extended DeepDDM, a pioneering approach using deep neural networks (DNN) to discretize subproblems from DDMs for solving partial differential equations (PDEs). By incorporating initial condition and ODE residual loss terms, it enhances optimization and accuracy. The method dynamically adjusts boundary terms using DNN, improving adaptability for complex PDEs. Its efficacy is demonstrated in modeling grid-forming inverters with a set of 13 differential equations, showing consistency with conventional DDM and reduced dependency on network architecture with increased overlapping size. Using an artificial neural network back-propagation method with Levenberg-Marquardt and Tansig activation, it achieves an R2 of 0.9973 with a training ratio of 0.8 and 20 epochs. These findings underscore the potential of DeepDDM in advancing SciML for solving complex multi-physics problems.

Name thesis:degree_name

Doctor of Philosophy

Discipline thesis:degree_discipline

Electrical Engineering

Grantor

University of Houston

Year dc:date.issued

2024

Author dc:creator

• Kalantari, Fatemeh

Advisors dc:contributor.advisor

• Krishnamoorthy, Harish Sarma
• Shi, Jian

Committee members dc:contributor.committeemember

• Viswanathan, Aparna
• Joardar, Biresh Kumar
• Fan, Lei

• Power systems and deep learning

Language dc:language.iso

English

Handle dc:identifier.uri

https://hdl.handle.net/10657/18315

OAI identifier oai:identifier

oai:uh-ir.tdl.org:10657/18315