Real-time and agile data-driven approaches enabling power grids to be smart
Loading...
Authors
Li, Wenting
Issue Date
2019-12
Type
Electronic thesis
Thesis
Thesis
Language
en_US
Keywords
Electrical engineering
Alternative Title
Abstract
Nowadays, the power system stability and reliability are challenged by the variable renewable energy and the frequent power outages. One promising remedy is to develop data-driven algorithms for power system intelligence, but practical deployment is not universal. One of the main concerns is the vulnerability of pure data-driven algorithms applied to variant power systems conditions. In this dissertation, we capture power system characteristics to develop online approaches that are robust to different power systems situations. We first want to identify different types of events through low-dimensional subspaces. We propose a data-driven realtime event identification method based on the measurements of Phasor Measurement Units (PMUs). The central idea is to characterize an event by the low-dimensional subspace spanned by the dominant singular vectors of a data matrix that contains spatial-temporal blocks of PMU data. The subspace representation is robust to initial system conditions and characterizes the system dynamics. A dictionary of subspaces that correspond to different events are established offline, and an event is identified online with the most similar event in the dictionary through subspace comparison. Numerical experiments on both simulated events and the recorded data validate the proposed method. Then we propose to identify successive events further to avoid cascading failures. Existing identification methods for single events may not accurately determine a subsequent event that occurs when the system is undergoing the disturbance of a previous event. We develop a data-driven event identification method that can accurately identify the types of overlapping events. Our approach only requires a small number of recorded PMU data of single events to train a two-layer convolutional neural network (CNN) classifier offline. We extract the dominant eigenvalues and singular values as features instead of training on time series directly. That reduces the required number of training datasets and enhances the robustness to measurement inaccuracy. We evaluate the method on simulated events in the IEEE 68-bus power system. Our classifier is demonstrated to be more accurate and stable than a direct application of CNN on time series. The robustness of the proposed method to the delay in event detection and noise is validated. Next, we consider locating faults in real-time to help power grids recover smoothly. Diverse fault types, fast reclosures, and complicated transient states after a fault event make real-time fault location in power grids challenging. Existing localization techniques in this area rely on simplistic assumptions, such as static loads, or require much higher sampling rates or total measurement availability. Unlike prior data-driven methods, the proposed classifier is based on features with physical interpretations that improve the robustness of the location performance. The accuracy of our CNN based localization tool is demonstrably superior to other machine learning classifiers in the literature. To further improve the location performance, a joint PMU placement strategy is proposed and validated against other methods. A significant aspect of our methodology is that under very low observability (7% of buses), the algorithm is still able to localize the faulted line to a small neighborhood with high probability. The performance of our scheme is validated through simulations of faults of various types in the IEEE 39-bus and 68-bus power systems under varying uncertain conditions, system observability, and measurement quality. Apart from monitoring the abnormal conditions, we improve the visibility of energy with the recorded data in the distribution system. We, for the first time, formulate the energy disaggregation at Substations (EDS) with partially labeled aggregate loads and develop an algorithm to separate individual loads in real-time. Unlike energy disaggregation at the household-level (EDH), the individual loads are not sufficient in EDS to learn patterns directly. Our approach based on a dictionary learning framework is modeless. In offline dictionary learning, we add column sparsity to the coefficients of the unlabeled loads by exploiting the partial labels. In online disaggregation, we decompose the testing aggregate time series into sparse linear combinations of sample training datasets. Both synthetic and realistic datasets validate our method. The proposed method outperforms the existing works in this EDS problem.
Description
December 2019
School of Engineering
School of Engineering
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY