Situational awareness in low-observable distribution grid: exploiting sparsity and multi-timescale data

Abstract

The power distribution grid is typically unobservable due to a lack of real-time measurements. While deploying more sensors can alleviate this issue, it also presents new challenges related to data aggregation and the underlying communication infrastructure. Limited real-time measurements hinders the distribution system state estimation (DSSE). DSSE involves estimation of the system states (i.e., voltage magnitude and voltage angle) based on available measurements and system model information. To cope with the unobservability issue, sparsity-based DSSE approaches allow us to recover system state information from a small number of measurements, provided the states of the distribution system exhibit sparsity. However, these approaches perform poorly in the presence of outliers in measurements and errors in system model information. In this dissertation, we first develop robust formulations of sparsity-based DSSE to deal with uncertainties in the system model and measurement data in a low-observable distribution grid. We also combine the advantages of two sparsity-based DSSE approaches to estimate grid states with high fidelity in low observability regions. In practical distribution systems, information from field sensors and meters are unevenly sampled at different time scales and could be lost during the transmission process. It is critical to effectively aggregate these information sources for DSSE as well as other tasks related to situational awareness. To address this challenge, the second part of this dissertation proposes a Bayesian framework for multi-timescale data aggregation and matrix completion-based state estimation. Specifically, the multi-scale time-series data aggregated from heterogeneous sources are reconciled using a multitask Gaussian process. The resulting consistent time-series alongwith the confidence bound on the imputations are fed into a Bayesian matrix completion method augmented with linearized power-flow constraints for accurate state estimation low-observable distribution system. We also develop a computationally efficient recursive Gaussian process approach that is capable of handling batch-wise or real-time measurements while leveraging the network connectivity information of the grid. To further enhance the scalability and accuracy, we develop neural network-based approaches (latent neural ordinary differential equation approach and stochastic neural differential equation with recurrent neural network approach) to aggregate irregular time-series data in the distribution grid. The stochastic neural differential equation and recurrent neural network also allows us to quantify the uncertainty in a holistic manner. Simulation results on the different IEEE unbalanced test systems illustrate the high fidelity of the Bayesian and neural network-based methods in aggregating multi-timescale measurements. Lastly, we develop phase, and outage awareness approaches for power distribution grid. In this regard, we first design a graph signal processing approach that identifies the phase labels in the presence of limited measurements and incorrect phase labeling. The second approach proposes a novel outage detector for identifying all outages in a reconfigurable distribution network. Simulation results on standard IEEE test systems reveal the potential of these methods to improve situational awareness.