UC Riverside

In the past few decades, the rapid development of the United States power system has led to the continuous expansion of transmission networks and an increasing number of phasor measurement units (PMUs) have been deployed on the power system. Although voltage and current phasor data can be obtained in a real-time operation environment, it is still challenging to effectively utilize PMU data in a large distributed system. Simply using off-the-shelf machine learning algorithms to process PMU data does not yield models with sufficient performance in practice. In this thesis, the physical dynamics of the U.S.power system was synergistically combined with machine learning to monitor and model a power transmission system.

The first aspect was real-time data-driven power system monitoring. We developed an efficient data-driven framework to detect voltage events from PMU data streams. In particular, we developed an innovative Proximal Bilateral Random Projection (PBRP) algorithm to quickly decompose a PMU data matrix into a low-rank matrix, a row-sparse event-pattern matrix, and a noise matrix. The row-sparse pattern matrix significantly distinguishes events from normal behavior. These matrices were then fed into a clustering algorithm to separate voltage events from normal operating conditions. Large-scale numerical study results on real-world PMU data show that the proposed algorithm achieved higher F1 and F2 scores with 50% less computation time.

The second aspect was to model dynamic electric power generator parameters. Accurate estimation of dynamic parameters is crucial to building a reliable model for dynamical studies and reliable operation of the U.S. power system. A physics-based neural ordinary differential equations (ODE) approach was developed to learn the generator dynamic model parameters using PMU data. We designed a physics-based neural network to represent the swing equations of the power system dynamics. The parameters of the generator dynamic model were iteratively updated using the neural ODEs and the adjoint method. By exploiting the mini-batch scheme in neural ODE training, the parameter estimation performance was significantly improved with more than 50% computation speed up.