Design and implement machine learning tools for gathering, analyzing, and visualizing diverse energy data sets; leading teams to address challenging energy systems problems.
Communicate
Communicate complex technical information through reports, websites, presentations, and videos for diverse audiences including scientists and engineers, energy professionals, and policy makers.
Educate
Educate undergraduate and graduate students in machine learning, programming, and project management skills through intensive applications and team-based learning experiences
Research
A selection of research projects from past and present.
Satellite Assessment
Machine learning to understand our world using remote sensing
Smart Meter Analytics
Applying machine learning to smart meter data
Wind & Solar Integration
Statistical modeling of alternatives for grid integration
Energy Storage
Optimizing storage operations to simulate market performance
Teaching
Below is a selection of courses taught, project teams led, and student mentorship experiences at Duke University. These experiences were offered across multiple departments and institutes including the Energy Initiative, Electrical and Computer Engineering, and the Masters of Interdisciplinary Data Science program.
Machine Learning
Intro to supervised and reinforcement learning
Python Programming
Short course for Python programming for data science
Bass Connections
Team-based learning through research project courses
Data+ and Climate+
Dynamic summer research project experiences for students
Selected Publications & Outputs
Selected publications and other works (e.g. datasets, videos) below. For complete list, please see CV
Automated Extraction of Energy Systems Information from Remotely Sensed Data: A Review and Analysis
PAPER
Applied Energy
Drones for Solar
Mapping Small Solar Home Systems Using Unmanned Aerial Vehicles and Deep Learning
PAPER
ISPRS
GridTracer
GridTracer: Automatic mapping of power grids using deep learning and overhead imagery
PAPER
JSTARS
PV & Building Assessment
Estimating Solar PV Capacity & Building Energy Use with Satellite Imagery
VIDEO
Building Energy
Estimating residential building energy consumption using overhead imagery
PAPER
Applied Energy
Energy Storage Optimization
Economic viability of energy storage systems based on price arbitrage potential in real-time US electricity markets
PAPER
Applied Energy
Synthetic Data for ML
The Synthinel-1 dataset: a collection of high resolution synthetic overhead imagery for building segmentation
PAPER
DATA
WACV
Solar PV Dataset
Distributed Solar Photovoltaic Array Location and Extent Dataset for Remote Sensing Object Identification
PAPER
DATA
Solar PV Segmentation
Semantic segmentation convolutional neural network for automatic detection of solar PV arrays in aerial imagery
PAPER
IEEE AIPR
Digital Transformations
How Data Science Can Enable the Evolution of Energy Systems
PAPER
Energy Data Analytics
How data science is enabling new ways of generating, transmitting, and consuming energy
VIDEO
Data Science Resources
How Data Science Can Enable the Evolution of Energy Systems
VIDEO
Contact / About
Kyle Bradbury
Director, Energy Data Analytics Lab
Primary Appointment: Assistant Research Professor, Electrical and Computer Engineering
Additional Roles:
Director, Energy Data Analytics Lab, Nicholas Institute for Energy, Environment, and Sustainability
Faculty member of the Masters of Interdisciplinary Data Science Program (MIDS)
Institution: Duke University
About: I lead applied research projects at the intersection of machine learning techniques and energy problems. My research includes developing techniques for automatically mapping global energy infrastructure and access from satellite imagery; transforming smart electric utility meter data into energy efficiency insights; and exploring the reliability and cost trade-offs of energy storage systems for integrating wind and solar power into the grid. I received both a Ph.D. in energy systems modeling and an M.S. in electrical and computer engineering from Duke University, as well as a B.S. in electrical engineering from Tufts University.
Education:
Doctor of Philosophy, Duke University, Energy Systems (Earth and Ocean Sciences Dept.), 2013
Master of Science, Duke University, Electrical Engineering, 2008
Bachelor of Science, Tufts University, Electrical Engineering, 2007
Machine learning for improving building energy efficiency
Plot of energy consumption in an example household versus the time of day. Individual appliance data is much more informative than aggregate building data.
For decades, only monthly electricity data was collected to provide a bill for building electricity consumption. However, the presence of smart meters has opened the door to huge quantities of data at sampling intervals ranging from 15 minutes down to 1 second. This project investigates techniques for decomposing the aggregate energy consumption from a building’s smart electric utility meters to provide device-level feedback (e.g. refrigerator, dishwasher, computer) and the time at which energy is consumed to improve building energy efficiency and operation.
Collaborators: Bohao Huang, Mary Knox, and Leslie Collins.
Energy Infrastructure Analysis using Remote Sensing Data
Using satellite imagery to better understand and plan energy systems globally
Energy systems are rapidly evolving in both the developed and developing world. Distributed energy resources such as solar photovoltaics are experiencing significant growth, but are not typically controlled or monitored by independent system operators of the grid. At the same time, one in five people around the world still lack access to modern electricity, and identifying optimal pathways to electrification requires detailed knowledge of local infrastructure and the precise locations of communities that would benefit most from electrification. Here, we work to fill these important information gaps by analyzing satellite imagery and other remotely sensed data using machine learning techniques, specifically through convolutional neural networks.
Using satellite imagery and machine learning, we are developing techniques to:
...Identify the location and capacity of solar photovoltaic arrays in the United States:
...Assess energy access in developing nations:
...Estimate building energy consumption:
...and map grid infrastructure including transmission and distribution lines to determine pathways for electrification in developing countries:
Collaborators: Jordan Malof (Pratt School of Engineering), Bohao Huang (Pratt School of Engineering), Wei (Wayne) Hu (formerly Energy Initiative), Leslie Collins (Pratt School of Engineering), Marc Jeuland (Sanford School of Public Policy), Bryan Bollinger (formerly Fuqua School of Business), Steve Sexton (Sanford School of Public Policy), T. Robert Fetter (formerly Nicholas Institute for Environmental Policy Solutions), Tim Johnson (Nicholas School of the Environment), Artem Streltsov (formerly Energy Initiative)
Energy Storage
Evaluating the arbitrage potential operation of energy storage systems in power markets
Figure showing a simulated energy storage device charge and discharge based on the real-time price of electricity
Energy storage systems (ESSs) can increase power system stability and efficiency, and facilitate integration of intermittent renewable energy, but deployment of ESSs will remain limited until they achieve an attractive internal rate of return (IRR). Linear optimization is used to find the ESS power and energy capacities that maximize the IRR when used to arbitrage 2008 electricity prices (the highest of the past decade) in seven real-time markets in the United States for 14 different ESS technologies.
Our results show that the profit-maximizing size (i.e. hours of energy storage) of an ESS is primarily determined by its technological characteristics (round-trip charge/discharge efficiency and self-discharge) and not market price volatility, which instead increases IRR. Most ESSs examined have an optimal size of 1–4 h of energy storage, though for pumped hydro and compressed air systems this size is 7–8 h. The latter ESSs already achieve IRRs >10%, but could be made even more profitable with minimal cost-reductions by reducing power capacity costs. The opposite holds for flywheels, electrical ESSs (e.g., capacitors) and a number of chemical ESSs (e.g., lead acid batteries). These could be made more profitable with minimal cost-reductions by reducing energy capacity costs.
Collaborators: Lincoln Pratson, Dalia Patino-Echeverri
Wind & Solar Integration
Comparing methods for integrating variable wind and solar energy systems into the grid
Intermittent energy resources, including wind and solar power, continue to be rapidly added to the generation fleet domestically and abroad. The variable power of these resources introduces new levels of stochasticity into electric interconnections that must be continuously balanced to maintain system reliability. Energy storage systems (ESSs) offer one potential option to compensate for the intermittency of renewables. Geospatially diverse siting of wind and solar, to make use of uncorrelated wind speeds and solar insolation, offer an alternative method for reducing the variability of renewable energy. The aim of this study is to explore the cost and reliability tradeoffs associated with using energy storage versus using geographically diverse siting for wind and solar integration.
Collaborators: Hyeongyul Roh, Artem Streltsov, Lincoln Pratson, Dalia Patino-Echeverri, and Xiaochen Sun
Bass Connections Team
Interdisciplinary student research project teams
Photo from the Pratt School of Engineering: "Data and the Duke Engineer"
Bass Connections is university-wide program for students to engage in interdisciplinary problem-based project teams. These students of multiple educational-levels spend a year working for credit on a team-based project guided by Duke faculty and staff.
Since 2013, I have led Bass Connections student research teams through projects to better operate or understand energy systems by using machine learning tools, resulting in the publication of datasets, award-winning student posters, and a generally a great team of students eager to take on the challenging energy problems of our world. Below are links to past project teams:
Collaborators: Jordan Malof, Leslie Collins, Timothy Johnson, Richard Newell
Data+ and Climate+
Student research mentorship program
Data+ and Climate+ are ten-week summer research experiences that welcomes Duke undergraduates and graduate student leaders. Students join small project teams, working alongside other teams in a communal environment. They learn how to marshal, analyze, and visualize data, while gaining broad exposure to the modern world of data science.
I have led a number of past project (listed below) that give students the experience of processing data to prepare for use in machine learning and through a series of training sessions help them to become familiar with the precepts of machine learning to apply to the data they curate through this summer research experience.
Data science is like painting. Your canvas is your dataset, your statistical techniques are your colors, and your technique and artistry comes through your programming. Data scientists with strong programming skills are able to quickly implement techniques and are more readily able to be more creative problem solvers and scientists. This online course is designed to will guide students through the basics of programming with Python with an eye towards the most important components needed to work with data. This course begins with an overview of the Python ecosystem and goes on to cover data types, operators, flow control, functions, modules, packages, objects, and classes. Since this course is data science focused, there is also a unit on computational programming that covers numpy, matplotlib (for plotting), and pandas (for working with data frames). There is also a discussion of best practices for scientific computing.
In almost every field, there is a need to draw inference from or make decisions based on data. The goal of this course is to provide an introduction to machine learning that is approachable to diverse disciplines and empowers students to become proficient in the foundational concepts and tools while working with interdisciplinary real-world data. Students learn to (a) structure a machine learning problem, (b) determine which algorithmic tools are applicable to a given problem, (c) apply those algorithmic tools to diverse, interdisciplinary data examples, (d) evaluate the performance of your solution, and (e) how to accurately interpret and communicate your results. This course is a fast-paced, applied introduction to machine learning that empowers students with the basic skills they will need in practice to both conduct analyses and effectively communicate their results to stakeholders.
