PNNL Centralizes AI Energy Research Under One Roof

January 12, 2024 by Shannon Cuthrell

The artificial intelligence center will coordinate research projects to improve the nation's energy security and grid reliability.

Power engineers use artificial intelligence (AI) and machine learning for applications ranging from smart grids to electric vehicles. To explore the potential of AI in all contexts, the U.S. Department of Energy (DOE) has established a research center in Washington State’s Pacific Northwest National Laboratory (PNNL) devoted to artificial intelligence projects.


Pacific Northwest National Laboratory researchers work on power grid projects.

Pacific Northwest National Laboratory researchers work on power grid projects. Image used courtesy of PNNL/by Andrea Starr


The Center for AI @PNNL will coordinate projects involving hundreds of researchers. Projects include developing AI tools to keep power flowing to homes and businesses and using machine learning to combine chemical compounds for lithium batteries in electric vehicles. Other research runs the gamut from training algorithms to detect theft in distribution circuits to modeling high-performance wind turbine designs

PNNL’s computer scientists have developed reinforcement learning techniques to evaluate the fundamentals of these applications before they are designed and developed. For example, one project focused on optimizing contextual two-stage stochastic programming, a method for decision-making under uncertain offshore wind farm market bidding. 

In 2023, PNNL’s AI-related teams published over 60 journal articles. The Center for AI @PNNL will advance those efforts and streamline researchers’ activities. 


Physics-Informed Machine Learning

The Center for AI @PNNL aims to boost the laboratory’s long-established specialty in physics-informed machine learning by using advanced algorithms based on Earth models, embedded systems, power grid components, and other sources. This method reduces the number of samples needed to train AI programs, making it useful in processing large datasets. Experiments in materials chemistry, for instance, could involve analyzing thousands of microscope images per second. 

Physics-informed machine learning projects often use digital twins to simulate batteries and inform the predictive development of grid-scale energy storage systems. 


PNNL’s digital twin technology produces a simulated flow battery.

PNNL’s digital twin technology produces a simulated flow battery. Images used courtesy of PNNL (screenshots at 3:18 and 3:23)


Some projects use exascale supercomputing, which employs advanced CPU and GPU infrastructure to enable billions of computations per second. A team developed a mathematical framework to study climate processes in subsystems of the DOE’s Energy Exascale Earth System Model, detecting rare events with smart sensor-equipped embedded systems. 

PNNL partners with the Exascale Computing Project (ECP), which works with companies to study emerging tech with commercial potential. The DOE’s exascale machines are among the world’s most powerful, performing a quintillion or more operations every second. One project developed a commercial-scale demonstration that predicts the performance of multiphase energy conversion devices to curb emissions at fossil fuel power plants. 

ECP’s exascale tools have also aided the development of small modular reactors, nuclear fission and fusion reactor materials design, and more areas. For example, predictive wind flow physics modeling helped harden wind plant design and layout against energy losses, allowing a higher wind energy penetration. 


Energy Resilience and Security Research 

Researchers at the Center for AI @PNNL will collaborate on training predictive AI models to secure the nation’s energy infrastructure. Here are a few recent projects in this field. 

  • Researchers generated synthetic data to train machine learning to detect theft in distribution circuits. Losses from illegal feeder tapping or meter tampering often remain undetected, and existing models rely on limited data since thefts are rare. PNNL researchers presented a co-simulation framework to generate reliable data to train algorithms for theft detection. 
  • An ECP project is developing algorithms to optimize grid response to weather events and other disruptions amid the transition to wind and solar resources. This addresses the national security challenge of optimizing grid dynamics under adverse conditions. 
  • PNNL’s DeepGrid open-source platform also uses deep reinforcement learning to help power system operators develop better emergency control protocols for their electric grids. 


Power Systems Research

AI tools allow researchers to analyze system behaviors, but the underlying models need quality datasets to train and test on. This will be a crucial focus for the Center for AI Research @PNNL. Two relevant examples include:

  • A PNNL team used a high-performance computing-based grid simulator, GridPACK, to recreate the voltage dynamics of a bulk power test system. The model captured fault-induced delayed voltage recovery, or FIDVR, a phenomenon that can be mitigated by an under-voltage load shedding (UVLS) control system. Researchers presented a dataset indicating voltage dynamics under different controls generated by UVLS methods and random noise. 

  • Researchers resolved underlying constraints in load frequency control systems for managing stable frequency in distributed multi-area power systems, which are increasingly transitioning from centralized to decentralized. They applied a new reinforcement learning-based output structured feedback framework to make a decentralized scheme considering limited information for each interconnected area’s generator control. They trained the model to simulate the dynamics of a physical grid. Tests confirmed that the feedback controller could control each area’s frequency while mitigating load uncertainty.