ML techniques on a satellite image
Applying visual object identification and machine learning techniques to satellite data can yield valuable energy insights.

Big data offers big opportunities to solve our world’s most daunting and complex energy challenges. That’s why Duke University’s interdisciplinary Energy Data Analytics Lab is developing groundbreaking applications of machine learning techniques and cultivating new data analytics talent for the sector.

Experts affiliated with the lab—including engineers, data scientists, and social scientists—are positioning Duke University as an international leader in the emerging area of energy data analytics. Among other projects, the lab has pioneered the application of visual object identification and machine learning techniques to satellite imagery for energy resource detection and mapping (scroll down to our Projects section for more information). One of the lab's long-term objectives is to create a map of global energy infrastructure that can be automatically updated.

Duke's Energy Data Analytics Lab is also creating a pipeline of talented innovators. Projects connected with Duke’s unique Data+ and Bass Connections programs accomplish lab research objectives while deepening undergraduate and graduate students’ research, project management, and communications skills. Thanks to grants from the Alfred P. Sloan Foundation, the lab launched programs for doctoral student fellows and postdoctoral fellows in fall 2018.

Founded in 2014, the Energy Data Analytics Lab is a collaborative effort of the Duke University Nicholas Institute for Energy, Environment & Sustainability (which houses it), the Rhodes Information Initiative at Duke (Rhodes iiD), and the Social Science Research Institute (SSRI).

Lab Leadership:

Events:

Fellowships:

 

Current Projects:

 

left:manual detection of solar panels,right:solar array detection on a satellite image by ML techniques
On the left you’ll see the location of solar panels in a neighborhood, as detected by humans painstakingly staring at satellite imagery. On the right, you’ll see what the solar array detection algorithm was able to detect (without anyone having to squint!).

Researchers at Duke have been pioneering the use of automated object identification and machine learning techniques to assess and map distributed generation, energy consumption, energy access, and other factors based on high-quality aerial imagery.

Researchers began by addressing a critical information gap regarding rooftop solar photovoltaic capacity and power generation: government statistical collections have focused on central station power, so information about distributed energy production at residential and commercial buildings is scarce. Project members first compiled, curated, and published a ground-truth data set, then trained image recognition algorithms to estimate the size and location of solar panels. Their methods can be used to improve solar PV estimates and aid government agencies and power grid independent system operators (ISOs) in evaluating the state of distributed PV deployment and use that information for planning purposes to increase system reliability and resilience.

Team members are now conducting related research applying automated object identification and machine learning techniques to assess and map energy access, building-level energy consumption, and infrastructure (e.g., network topologies of electric power lines).

One of the project’s long-term objectives is to create a map of global energy infrastructure that can be automatically updated, and work is well underway on this front. The team’s work could inform other researchers’ efforts to tackle a range of topics beyond energy, identifying and measuring nearly any physical object that’s visible from above— like monitoring, assessing, managing, and predicting urban development or changes in agricultural production.

Researchers: Jordan Malof (Pratt School of Engineering), Bohao Huang (Pratt School of Engineering), Kyle Bradbury (Managing Director of Energy Data Analytics Lab, Nicholas Institute), Leslie Collins (Pratt School of Engineering), Marc Jeuland (Sanford School of Public Policy), Bryan Bollinger (Fuqua School of Business), Steve Sexton (Sanford School of Public Policy), T. Robert Fetter (Duke Sanford Center for International Development), Tim Johnson (Nicholas School of the Environment) .

Related student-faculty projects:


mapping electricity access for Bihar,India

More than 15% of humanity has no access to electricity, and far more have access only to intermittent supplies. At Duke and elsewhere, researchers seek to better understand the drivers and impacts of electrification on health, land use, the environment and the local economy. Yet current methods for assessing access rely on household surveys or highly aggregated (e.g., national-level) data sources.

This project focuses on evaluating electricity access in developing countries through machine learning techniques applied to aerial imagery data.

The project launched with compilation, curation, and publication of an initial ground-truth data set for 36,000 villages in the Indian state of Bihar, and also collected measurements relevant to electricity consumption including lights at night data and irrigation metrics. Project team members created an Amazon MTurk tool that crowdsourced the annotation of key electricity indicators (such as power plants and transmission lines) in imagery data.

Team members have worked to pilot and apply automated algorithms for generating spatially-disaggregated data on electricity access in developing countries using aerial imagery. This work has been supported by Bass Connections, a Catalyst Grant from the Nicholas Institute for Energy, Environment & Sustainability, and an Energy Research Seed Fund grant

Researchers: Jordan Malof (Pratt School of Engineering), Kyle Bradbury (Managing Director of Energy Data Analytics Lab, Energy Initiative), Marc Jeuland (Sanford School of Public Policy) Leslie Collins (Pratt School of Engineering), T. Robert Fetter (Duke Sanford Center for International Development), Robyn Meeks (Sanford School of Public Policy)

Related student-faculty projects:

 

Past Projects: 

graph,REDD aggregate laod with appliance subset
Examples of individual appliance data disaggregated from aggregate building data

Electric meter data, advanced thermostats and other components of modern smart grid and systems connected in an "internet-of-things" have the potential to enable significant insights and energy automation in buildings. The Energy Data Analytics Lab is exploring methods and applications for non-intrusive load monitoring (NILM), which breaks down aggregate energy consumption data from a building’s smart electric meter to provide feedback on each type of device that is consuming energy.

These techniques may allow building owners to cheaply automate building energy audits, identify energy efficiency improvements, predict equipment failure, and maximize cost savings by using less power or using it at a time when the cost is lower. Researchers are looking at ways to improve NILM techniques and applications, and how data from the growing number of connected devices in a building may be leveraged to optimize building energy consumption through automation.

Researchers are exploring the impact of the sampling rate of smart meters on NILM performance to better assess what smart meter hardware infrastructure is needed to enable energy efficiency and operational improvements from NILM.

Researchers: Mary Knox (Pratt School of Engineering), Bohao Huang (Pratt School of Engineering), Leslie Collins (Pratt School of Engineering), Kyle Bradbury (Managing Director of the Energy Data Analytics Lab, Energy Initiative), Richard Newell (in his former role as Energy Initiative Director)

 

Related student-faculty projects:

Plot,difference in productivity between conventional & unconventional wells
Plot demonstrating the difference in productivity between conventional and unconventional wells

Horizontal drilling and hydraulic fracturing have fundamentally changed the oil and gas industry, with a significant impact on natural gas markets and pricing. Through these manufacturing processes, gas can be quickly and reliably extracted from shale formations. This suggests that natural gas production from shale wells should be more responsive to natural gas prices, which should increase the elasticity of gas supply and dampen gas price volatility. A team of Duke researchers is using a large dataset on oil and gas wells to investigate the changing nature of the gas supply curve. The data provided by DrillingInfo contains detailed information on hundreds of thousands of oil and gas wells in the United States, and tens of millions of well-month observations. The research will also be expanded to explore tight oil production.

Researchers: Brian Prest (Nicholas School of the Environment), Richard Newell (Former Director, Energy Initiative)

graph,energy-saving impact of automation over behaviorally-motivated changes
Demonstrates the energy-saving impact of automation (green line) over behaviorally-motivated changes

While at Duke, former public policy professor Matthew Harding worked to compare how smart algorithms and devices may work better than trying to change individual behavior when it comes to improving energy consumption decisions. As utilities and appliance manufacturers offer advanced data on their products, organizations and individuals can use the information to increase energy efficiency and reduce costs by relying more on automated systems that respond to system signals such as weather. Harding was looking in particular at how consumers respond to incentives to improve energy conservation or adopt green power.

Researchers: Matthew Harding (former faculty member, Sanford School of Public Policy)

Water heating in the residential building sector is the second-highest energy end use after space heating. Modern water heaters come in many forms ranging from conventional storage water heaters to tankless and heat pump water heaters. Although many energy efficiency programs advocate the replacement of conventional systems with heat pump water heaters, conventional tanks still dominate the market and are likely to be around for years to come. Using state-of-the-art sensor data from industry partners, we are developing control systems to maximize the energy efficiency of these systems, evaluating the potential energy and cost savings to water heater owners.

Researchers: Mengyang Lin (Former Research Analyst, Energy Initiative), Kyle Bradbury (Managing Director of Energy Data Analytics Lab, Energy Initiative), Richard Newell (Former Director, Energy Initiative)

Contact: 

  • Business and policy leaders who wish to discuss potential projects with the Energy Data Analytics Lab team should contact Dr. Kyle Bradbury, Managing Director: Kyle.Bradbury@Duke.edu or (919) 613-2411.
  • Journalists seeking to interview a Duke University energy data expert should contact Braden Welborn, Director of Communications: Braden.Welborn@Duke.edu or (919) 613-0436.

Mailing Address

Nicholas Institute for Energy, Environment & Sustainability
Box 90467
Durham, NC 27708

Street / Delivery Address

Nicholas Institute for Energy, Environment & Sustainability
140 Science Drive
Gross Hall, Suite 101
Durham, NC 27708

919-613-1305 

 

nicholasinstitute@duke.edu

  • Kyle Bradbury
    Assistant Research Professor, Pratt School of Engineering & Managing Director, Energy Data Analytics Lab

    Kyle brings experience in machine learning and statistical modeling to energy problems.