Artificial Intelligence for the Power Grid

Vision models allow researchers to quickly understand electrical assets at a specific site using high-resolution imagery collected by drones, ground-based cameras, and other satellite and aerial imagery. AI vision models process the imagery; capture metadata about capacity, connectivity, and potential signs of physical wear; and ultimately accelerate digitalization and analysis. User-verified AI processing can speed up grid digitalization and asset assessment and reduce costs by not requiring physical checks of each asset.

NLR researchers are using vision models to assess national grid expansion potential, quickly process single-line diagrams, and assist utilities with asset mapping and the development of grid digital twins.

Foundation models are large, reusable AI models trained on diverse datasets so they can support multiple downstream tasks rather than a single narrow application. For energy systems and grid research, foundation models could integrate knowledge from operational data, sensor measurements, weather and load data, equipment information, simulations, and other sources. Foundation models can be used and fine-tuned for applications such as forecasting, equipment and anomaly detection, grid-state assessment, scenario generation, control and operations support, and analysis of large simulation datasets.

As part of the U.S. Department of Energy (DOE) Genesis Mission's Transformational AI Models Consortium, NLR is collaborating with other national laboratories and partners to advance foundation models for the electric grid. These models can learn from a variety of grid data and topologies, including power-flow snapshots and weather-driven operating or network-topology scenarios, to capture relationships that reflect the physics and behavior of the grid. Foundational models can permit researchers to evaluate many more grid conditions than is practical with conventional simulations alone.

Many portions of the electric grid respond to grid disruptions without immediate human intervention, automatically diverting power around problems or triggering protection devices if a surge in current is detected. Understanding how grid components and controls respond across a range of disruption scenarios can help planners and operators identify vulnerabilities, compare mitigation strategies, and prioritize investments.

However, repeatedly running detailed power-flow, dynamic, or electromagnetic-transient simulations across thousands of possible scenarios can be prohibitively time-consuming. NLR is using AI-powered surrogate models to learn statistical relationships among operating conditions, disturbances, control actions, and system outcomes using time-series measurements and physics-based simulation results.

These surrogate models can rapidly estimate outcomes such as voltage, power flow, equipment loading, stability margins, outage risk, and recovery performance. By screening a much larger set of scenarios in less time, surrogate models can support uncertainty analysis, faster planning studies, and near real-time decision support, while high-fidelity simulations and human review remain part of the validation process.

NLR researchers are developing AI agents that can work in conjunction with the laboratory's power-flow and grid planning models. These agentic workflows can help users define a study, select informed assumptions, prepare model inputs, run simulations, check results, and summarize findings through a conversational interface using the stakeholder’s own grid data. The agentic workflows can enable users to, for example, chat with agents to execute a grid planning study using informed assumptions based on the stakeholder's own grid data.

Users including utilities, grid planners, modelers, operators, regulators, and program managers can use agentic workflows to reduce time spent writing and debugging scripts, speed up comparison of alternative scenarios, produce more consistent and reproducible workflows, and have broader access to rigorous grid analysis. Human experts remain responsible for reviewing assumptions, validating outputs, and approving decisions.

Large datasets can contain a wealth of valuable information but identifying insights in a way that is cost effective and timely can be a challenge. Researchers are using AI to process and organize big data to inform new research directions, technical assistance, and grid decision-making. NLR develops tools that harness the power of large language models to automate the compilation and continued maintenance of datasets, such as an inventory of state and local codes and ordinances pertaining to energy infrastructure.

Generative AI, a subset of AI tools, can be trained to deliver reliable information and decision-making support across a range of possible applications within power systems.

Generative AI can be used to help solve grand challenges in the power sector by:

• Realizing proactive, real-time energy system operations
  Generative AI, in concert with next-generation AI foundation models, can revolutionize grid operations by providing proactive decision support and predictive online control to improve efficiency, reliability, and resilience.

• Building cyber- and all-hazards resilient and secure energy systems
  Harnessing AI provides a strong opportunity to achieve a cyber-resilient and all-hazards resilient grid by reducing blackouts and brownouts and ensuring that all communities have access to affordable, reliable, and clean electricity.

• Designing and planning an affordable and reliable electricity system by 2035
  AI will help change the planning paradigm for the future power grid by providing fast and efficient models, high-fidelity scenarios, and stochastic optimization schemes for large-scale integrated energy systems.

Read more about generative AI for grid modeling in eGridGPT: Trustworthy AI in the Control Room, NLR Technical Report (2024).