Distributed energy resources (DERs), such as distributed solar photovoltaic (PV) systems, electric vehicles, distributed energy storage, and demand response, are being deployed in the power distribution system at an unprecedented pace. Though DERs bring environmental and technological benefits, challenging technical problems arise as well. Distribution networks must be actively managed and planned/upgraded accordingly to accommodate DERs and coordinate their operations. All of these depend on the solving of the technical problems of accurate phase identification, network parameter estimation, DER adoption prediction, and long-term load forecasting in the distribution system. In this dissertation, I use machine learning and data analytic techniques to address these challenging problems, which are critical to the adoption of DERs.
To address the problem of phase identification, we study multiple methodology approaches. Two unsupervised learning algorithms are developed based on smart meter data and supervisory control and data acquisition (SCADA) data. The first algorithm leverages linear dimension reduction and centroid-based clustering. The second algorithm further improves the phase identification accuracy by nonlinear dimension reduction and density-based clustering. In the third approach, a maximum marginal likelihood estimation approach based on physics-informed model is proposed, which is physically interpretable and more accurate.
To address the problem of three-phase network parameter estimation, we develop a maximum likelihood estimation approach based on a physics-informed model to estimate the serial impedance of three-phase lines. A more advanced method based on graphical learning model is then developed to provide more accurate parameter estimation.
To address the problem of DER adoption prediction, we study the adoption of distributed commercial solar PV systems by developing a generalized Bass diffusion model. This model is capable of forecasting solar PV adoptions and quantifies the impact of solar PV costs and government incentive programs on the adoption.
To address the problem of long-term load forecasting, we develop comprehensive models based on supervised learning to forecast the diversity factor (DF) of distribution feeders at high accuracy. We also quantifies the importance of different influential factors and analyze how they affect DF.
