“By leveraging machine learning, we can help utility companies be more proactive instead of reactive.”
– Brenda Ng
The U.S. power distribution system is a complex, interconnected network of varied functional components. Ensuring that customers have a continuous supply of electricity requires sophisticated monitoring of these specialized components. Every part of the network—generators, transmission lines, substations, transformers, and distributions lines—could experience failures due to aging infrastructure, human-made reasons, or natural causes like storm damage or overgrown vegetation. The Department of Energy’s Grid Modernization Laboratory Consortium is tackling such power grid challenges through multi-year research efforts spearheaded by the national laboratories.
One such effort is led by LLNL as part of its Cyber and Infrastructure Resilience (CIR) portfolio. In this project, machine learning (ML) is used to improve situation awareness of the energy grid—in particular, to identify critical failures before they happen. Team lead Brenda Ng states, “By leveraging machine learning, we can help utility companies be more proactive instead of reactive.”
The team’s goals are to create an operationalized framework that detects failures at different levels within the power grid, with focus on wildfire prevention as the primary use case. The three-year “Incipient Failure Identification for Common Grid Asset Classes” project is a collaboration with Oak Ridge National Laboratory, National Energy Technology Laboratory, and Sandia National Laboratories as well as academic (Texas A&M and University of Pittsburgh) and industry partners (Pacific Gas and Electric Company, Electric Power Board of Chattanooga, National Rural Electric Cooperative Association, Electric Power Research Institute, OSISoft, and Corning).
The heart of the project involves collecting and analyzing a variety of data from multiple points in the network. For example, acoustics and vibrations from substation components that signal or precede potential failures can be detected with fiber optic sensors. Consumer-level data can be captured via smart meters, and historical databases can yield contextual information on past failures. By using ML to extract patterns and combining these patterns from different sources to improve situation awareness, the team aims to enable earlier detection of incipient faults and provide a transferable approach for wider adoption.
According to Ng, the first step is to learn meaningful representations of this massive, multimodal data. She explains, “We will be using deep learning to extract relevant features from each data stream, then combine all of these features to obtain a holistic glimpse of the health of the grid components.” The combined features will be used to predict the health and degradation of energy grid assets via Gaussian processes, which provide uncertainty estimates necessary for risk assessment.
Many similarities exist between the energy grid and the human body. In fact, much of the analytics used for this project is adopted from temporal multimodal learning techniques developed for healthcare, as part of a Laboratory Directed Research and Development (LDRD) project headed by Priyadip Ray. Ray’s LDRD team applies deep learning methods to uncover meaningful information from patients’ vitals, lab tests, medication histories, and visit logs to improve understanding of prognostic factors for a given disease—thus enabling prediction of the risks of adverse outcomes from the prescribed course of treatment. Algorithms that detect symptoms of pending failure or estimate an underlying health status have direct analogs with respect to the energy grid. “People and processes generate massive amounts of data,” says Ng. “Let’s take advantage of that information to improve everyday operations and, in this case, improve the reliability and resiliency of the energy delivery system.”
With its enormous scale and myriad of data modalities, the power grid is an untapped goldmine for data scientists. Ng advises, “Examine your pipeline. Find its weak points and opportunities for automation. See if data science methods can help you improve a problem and thus become a force multiplier.”
LLNL team members on the project are Brenda Ng, Hannah Burroughs, and Ryan Chan. The associate program leader for CIR is Emma Stewart.