NREL's artificial intelligence (AI) research targets technology advancement through machine learning (ML), reinforcement learning (RL), multidisciplinary deep learning, and an enabling high-performance computing (HPC) infrastructure.
NREL is seeing success with generalized AI, motivated by applications from autonomous vehicles and systems to machine-guided inverse design.
We use AI, along with ML, as an ideal tool for deriving new insights from analysis of very large data sets. At NREL, an ever-growing emphasis on underlying, state-of-the-art mathematics underpins AI, enabling new applications. Applied math research is necessary either to adapt ML techniques to energy domain data or to adapt the problem formulation to take advantage of mathematical structure.
We have successfully used ML and AI for applied predictive and explanatory modeling. In these cases, we used existing data to predict a future state of a system or unknown quantity.
In control—especially of large energy systems that challenge traditional methodology—we use the technique of RL. We have explored, among others, the use of RL for wind farm control, grid-interactive building control, connected and autonomous vehicles control, smart home control (e.g., HVAC and water heaters), and transportation fleet management.
Multidisciplinary Deep Learning
NREL has a multidisciplinary team of researchers collaborating on deep learning efforts. One robust effort is underway to create and maintain software examples for leveraging parallelism and HPC in ML and AI training workflows.
We also regularly facilitate deep learning training sessions, which allow NREL cohorts to engage in professional development together, often well beyond the initial activity's scope.
High-Performance Computing Infrastructure
NREL fuses data streams, modeling and simulation, and ML into its HPC workflows, a paradigm anchored in its data centers. Our Stratus service's multilayered capabilities make cloud services accessible to the lab-based scientific community, while on-premise cloud resources provide flexible data-centric capabilities.
To align optimal capability, data, and user intent in the HPC workflow, we consider:
- Sensor data intake, cleaning, and transforming edge cloud suitability
- Synthetic data generation through modeling and simulation, computationally heavy ML training, and hyperparameter optimization for HPC suitability
- Cloud interference and deployment.
NREL uses a deep learning technique (super-resolution) with a novel ML approach (adversarial training) to enhance the spatial and temporal resolutions of global climate data with greater accuracy and speed. These high-resolution climate data are important for predicting variations in wind, clouds, rain, and sea currents that fuel renewable energies and inform future climate scenarios.
Athena is a digital twin model of the Dallas-Fort Worth airport that leverages data-driven statistical modeling and AI to simulate the impacts of future capacity expansion scenarios. This is a collaborative effort funded by the U.S. Department of Energy Vehicle Technologies Office and industry and is led by NREL in partnership with Oak Ridge National Laboratory and Dallas-Fort Worth International Airport.
Deep Reinforcement Learning Based Volt-VAR Optimization in Smart Distribution Systems, IEEE Transactions on Smart Grid (2021)
Deep Reinforcement Learning for Dynamic Control of Fuel Injection Timing in Multi-Pulse Compression Ignition Engines, International Journal of Engine Research (2021)
A Data-Driven Operational Model for Traffic at the Dallas Fort Worth International Airport, Journal of Air Transport Management (2021)
Proof-of-Concept of a Reinforcement Learning Framework for Wind Farm Energy Capture Maximization in Time-Varying Wind, Journal of Renewable and Sustainable Energy (2021)
Deep Reinforcement Learning for Automatic Generation Control of Wind Farms, American Control Conference (2021)
Artificial Intelligence and Critical Systems: From Hype to Reality, Computer (2020)
Machine Learning of Combustion LES Models from Reacting Direct Numerical Simulation, Data Analysis for Direct Numerical Simulations of Turbulent Combustion (2020)
A Methodology for Reliability Assessment and Prognosis of Bearing Axial Cracking in Wind Turbine Gearboxes, Renewable and Sustainable Energy Reviews (2020)
A Distributed Reinforcement Learning Yaw Control Approach for Wind Farm Energy Capture Maximization, American Control Conference (2020)
Heterogeneous Machine Learning on High Performance Computing for End to End Driving of Autonomous Vehicles, WCX SAE World Congress Experience (2020)
Deep Learning for Presumed Probability Density Function Models, Combustion and Flame (2019)
Message-Passing Neural Networks for High-Throughput Polymer Screening, Journal of Chemical Physics (2019)
Distributed Reinforcement Learning with ADMM-RL, American Control Conference (2019)
An Open Experimental Database for Exploring Inorganic Materials, Scientific Data (2018)
Prediction and Characterization of Application Power Use in a High-Performance Computing Environment, Statistical Analysis and Data Mining (2017)