Data Driven State Estimation in Distribution Systems
Increasing penetration of distributed energy resources (DERs) and electric vehicle (EV) charging load has brought intermittency into the distribution systems. Specially, power generation from the onsite renewable energy sources such as solar photo-voltaic is known to be intermittent. Together with variable EV charging loads that are highly dependent on user behavior, bi-directional power flows can be introduced into the distribution networks that results in unclear power flow direction, which may violate grid operating constraints such as voltage constraints and current magnitude constraints. Hence, it is necessary to study and implement distribution system state estimation (DSSE) as it provides voltage, current, and power information of every single element in the distribution system, and further enables examination of the operating constraints that help maintain the system in a normal operating status.
Unlike transmission systems that are well-monitored with measurement redundancy, distribution systems are usually characterized by measurement scarcity and rely on pseudo-measurements modeled from historical data. Traditional pseudo-measurement methods assume static load profiles and fail to address the impact of DERs and EV charging penetration as a dynamic process. This dissertation proposed a pseudo-measurement modeling method at phase-level using user-level data to model the change of load profiles due to this penetration. Numerical analysis using real-world EV charging load, residential load, commercial load, and solar generation data demonstrates the merits of the proposed pseudo-measurement modeling method.
Moreover, state estimation is aimed at enabling real-time applications for grid operations and therefore a fast runtime is desirable. However, distribution systems also have unbalanced phases and a larger network size compared to their transmission counterparts, resulting in undesirable runtime efficiency. In this dissertation, a two-step DSSE framework is proposed. First, a surrogate model is developed to generate a fast-yet-coarse prediction of the system states. Second, this state surrogate is further fed into the state estimator as the initial state values to improve the state estimation runtime efficiency. Deep neural networks (DNNs) and long short-term memory (LSTM) networks are proposed to model the states, and autoencoder networks are further introduced to compress the input so that the proposed networks are trainable on large distribution systems. The proposed methods are tested on IEEE 123-bus system and showed that the runtime of DSSE is reduced by 73%. The proposed methods are further tested on IEEE 8500-node benchmark system with an 83% DSSE runtime reduction.
Finally, this dissertation introduces the development of a measurement system infrastructure with power quality analyzer (PQA) and micro phasor measurement unit (PMU) systems. The data collected are used in numerical studies mentioned above. We further introduce a monitoring application based on the measurement infrastructure and conclude the lessons learned from the implementation of the systems.