Semi-Supervised Deep Learning Frameworks for Transmission-Scale Load Disaggregation and Behind-the-meter Solar Prediction

While bringing environmental benefits, the proliferation of renewable energy resources poses challenges to transmission system operators due to their volatility nature. Thus, it is essential for Regional Transmission Operators (RTOs) to accurately extract load profiles for nodes with substantial BTM solar injection. This dissertation presents novel transmission-scale load disaggregation and BTM solar prediction frameworks, addressing lack of visibility for and enhancing situational awareness for transmission operators. Unlike distribution-level BTM solar generation which has ground-truth data, transmission-level BTM solar generation lacks such visibility. To address aforementioned challenge, spatial and temporal relationships between nodal and zonal load profiles are proposed and validated. Validated relations and used to disaggregate transmission-level load profiles. A proxy solar profile within each zone is utilized to segment the dataset. Finally, a semi-supervised model is developed to disaggregate nodal load demand profiles. To evaluate the outcomes without ground truth, cross-zero points and various distance matrices, including Wasserstein, symmetrical KL, and area difference metrics are adopted. Disaggregation models utilizing Linear, bi-linear, and non-linear features are validated with real world data from PJM Interconnection, respectively. Based on disaggregation framework, a self-supervised, transmission-scale BTM solar prediction framework is developed based on Timeseries Dense Encoder (TiDE) algorithm, emphasizing low computational costs and high accuracy for large datasets. This work presents a comprehensive, bottom-up framework for disaggregating transmission-scale load profiles and predicting BTM solar generation.