While sensing is becoming more prevalent in power systems, electric utilities still often lack an accurate real-time picture of the behavior of distributed energy resources, such as electric loads and distributed solar power. Such information would help system operators, utilities, energy efficiency providers, and demand response providers improve power system reliability, economic efficiency, and environmental impact. However, sensing infrastructure is costly, especially when considering the large number of quantities we might be interested in measuring. The goal of this research is to develop methods to infer the real-time behavior of aggregations of distributed energy resources from existing power system measurements, which are hierarchical, heterogeneous, incomplete, and of varying quality. To do this, the researchers are applying and extending emerging online learning techniques that leverage dynamical system models. While methodological developments are grounded in the power system application at hand, the extensions are informing new research directions for signal processing. Knowledge of what can and cannot be inferred from existing data will help utilities determine the value of additional sensors, what type of sensors are needed for different applications, and where to put them. This will also help policy makers determine which infrastructure investments are worthwhile and the need, if any, for subsidies. Additionally, the results will inform energy policy discussions on the value and cost of consumer privacy, which will help develop policies that better balance the objectives of power system operators, utilities, and third-party companies with those of consumers.
The research is applying an emerging technique, online learning with dynamics (OLWD), to determine what can and cannot be inferred from both existing power system measurements and measurements that we might expect to have in the near term. Contemporary online learning algorithms do not handle time-varying phenomena because they do not include dynamical models, and classical online estimation algorithms are not robust to model misspecification. In contrast, OLWD uses a collection of models (of arbitrary form) and the algorithm simultaneously estimates state and selects the model or combination of models that best predicts the state at the next time step. OLWD is based on one of the most successful current online optimization algorithms, inheriting many of its appealing properties. While the approach is well-suited to the problem, the theory is incomplete. A key component of the research is extend OLWD to handle measurements that are hierarchical, heterogeneous, incomplete, and of varying quality. The researchers are exploring both passive online inference and active online inference, where the latter uses external control (e.g., of controllable loads and curtailable solar photovoltaics) to enhance learning. Additionally, the researchers are characterizing trade-offs between system cost, inference accuracy, and consumer privacy.
