Learning and Control Applied to Demand Response and Electricity Distribution Networks

Abstract

Balancing the supply and demand of electrical energy in real-time is a core task in power system operation. Traditionally, this balance has been achieved by controlling power plants, but increasing amounts of renewable energy generation increases the variability in generation and requires additional energy balancing capacity. An alternative to providing this additional capacity via power plants is to provide signals to loads that induce changes in their demand, which is referred to as demand response.

There exists a large potential capacity for demand response using residential loads, but enabling these loads to participate in demand response requires communication and sensing capabilities. Thermostatically controlled loads (TCLs) are ubiquitous in residences and have inherent flexibility as they cycle on and off during normal operation. Coordinating on/off switching of TCL aggregations can provide energy balancing. However, TCLs are a spatially distributed resource that require sensing and communication infrastructure to enable demand response capabilities. A key to realizing cost effective residential demand response is minimizing infrastructure costs while maximizing the accuracy of the provided energy balancing, which results in increased revenue while improving reliability in the power system.

The main contribution of this dissertation is to show that advanced algorithms can leverage existing infrastructure to make energy balancing with loads feasible in the near-term, which improves the reliability, economics, and environmental impact of the power grid. The dissertation first presents control algorithms, estimation algorithms, and models for residential demand response on fast timescales, i.e., on the order of seconds. Following this, the dissertation presents online learning algorithms for real-time feeder-level energy disaggregation within an electricity distribution network, which can be used to estimate the demand-responsive load in real-time. Methods for both topics are developed to operate within the capabilities of existing communication and sensing infrastructure, which reduces the implementation costs of the methods.

Control and estimation algorithms are developed that address communication delays while taking into account realistic measurement availability. Results indicate that incorporating delay information into the algorithms can mitigate the effects of communication delays, allowing demand response providers to reduce infrastructure costs by using less expensive, lower quality communication networks. Additional work adapts three existing residential demand response models for a more detailed simulation environment, modifies each model to be more accurate in this environment, and benchmarks the model variations against each other. Results indicate that the model modifications produce more accurate predictions versus the unmodified models. Improving modeling accuracy can improve the reliability of the system and increase revenues for a demand response provider by improving the performance of model-based control and estimation algorithms.

The energy disaggregation algorithms seek to separate the measured demand of a distribution feeder into components (e.g., the demand-responsive load and the remaining load) as feeder-level measurements become available. An online learning algorithm is adapted to perform real-time energy disaggregation using active power measurements of the total demand on the distribution feeder. Results indicate that the algorithm is able to effectively separate the air conditioning demand from the remaining demand connected to a distribution feeder. This algorithm is then extended to include reactive power, voltage, and smart meter measurements. Results indicate that the availability of additional real-time measurements leads to more accurate disaggregation of the demand components. Additional work in state estimation establishes connections between the online learning methods used and Kalman filtering algorithms.