Recent developments in behavioral science, machine learning, and information and communication technologies are fundamentally transforming the questions and methods that can be applied to sustainability science. Research in decision making is transitioning from rational expectations towards bounded rationality models, machine learning excels at prediction rather than hypothesis testing, and ubiquitous sensor networks can elucidate new insights regarding previously unobserved system dynamics. This dissertation combines these three approaches to explore demand side features and strategies for sustainable decarbonization through various case studies of global, national and urban energy systems.
Behavioral science provided a strong theoretical framework with which to frame many of the insights that I encountered through field-research and data-mining. Why and how is it that some relatively poor countries, with weaker traditional notions of institutional strength make significant progress towards decarbonization? Why do relatively wealthier countries, with supposedly stronger institutions, and with abundant renewable resources fail to make such progress? Why do low-income households fail to invest their savings from energy efficiency or long-term energy efficiency retrofits? Why would a low-income household prefer high-resolution information over cash? How can we design more effective support mechanisms for sustainability at the national and household level? Ryan and Deci’s insights on intrinsic motivation (and all the subsequent work by other researchers they’ve inspired) (1, 2), Kahneman and Tversky’s work on prospect theory, the endowment effect, and loss aversion (3–5), Thaler’s insights on nudging and savings (4, 6), and Mullainathan and Shafir’s work on the psychology of scarcity (7, 8), have all provided rich insights with which I address many of these questions.
This dissertation also places a strong emphasis on data mining and statistical inference. Data mining and machine learning approaches are appealing because they are theory agnostic, can deal with nonlinearities, and encourage the researcher to collect as much data as possible. While this dissertation makes no contributions to improving the accuracy of data mining and machine learning techniques, it does provide new ways of bringing data together for the purpose of sustainability science. Rather than using causal inference to explore a specific hypothesis, I organize disparate ‘long and wide’ data and use data mining and prediction approaches (e.g., principal component regressions, nearest neighbors, random forests) to extract the features that can best predict my dependent y variable. In some cases, I also use Bayesian inference to characterize the size and uncertainty of the outcomes measured in my field-work, as well as to build posterior distributions from several seemingly disparate data streams. Bayes, philosophically, is particularly interesting as its usefulness increases as more data is collected, encouraging the researcher to continuously collect more evidence in order to test the strength and uncertainty of an initial hypothesis. A strong hypothesis would need a lot of data to be refuted, and thus, one has to collect an extraordinary amount of evidence to change a well-established norm. On the other hand, a weak hypothesis that is not backed up by data can be easily refuted. More importantly, Bayes lends itself to the evaluation of strong theories. Strong theories estimate the magnitude of parameter values and their credibility, not merely reject null values (9, 10). I use Bayes, data-mining and machine learning approaches in my work as they provide a flexible yet rigorous analytical tool box for which to extract meaning from data.
The rapid cost reduction in sensor networks suggests that in the near future they will be the bridge between behavioral science and data mining. While they have been used in manufacturing and industry for decades, they are now proving to be fundamental for research at the intersection of behavior, technology and sustainability. For example, sensors can be used to effectively monitor the efficacy and usage of cook stoves, water filters, and water delivery in field trials (11), to validate user responses in surveys and interviews against sensor data, and to evaluate whether or not a socio- technical intervention geared towards sustainability is being used appropriately. While the gold- standard for field-technology trials is randomized controlled trials (RCTs), the evaluation of an intervention is only valid if it was implemented without error; if not, what the RCT is truly evaluating is the quality of the intervention rather than the research questions themselves. Sensors allow researchers not only to monitor and troubleshoot, but also to uncover hidden insights about a system – how do rainfall and temperature affect appliance usage? What environmental factors affect the quality of the intervention? How can the technology be changed in the future to address usage issues related to environmental and behavioral factors? Similarly, information and communication technologies (ICTs) allow for a cost-effective two-way communication pathway with users: nudging, questions, complaints, AB testing, and feedback become immediately available with ICTs.
How are these three broad and overarching themes related to sustainable decarbonization? Sustainable decarbonization, here, is defined as the equitable reduction in total energy demand, accompanied by the increased adoption and use of low-carbon life styles and technologies. Equity in this context, is defined as equal opportunity of access to these lifestyles and technologies, while in consideration of barriers to entry (e.g., gentrification, racism, affordability, income), and the design and implementation of mechanisms to overcome them. At the global and national-level, equitable reductions in energy consumption and increase in the growth of low-carbon technologies will not occur without contextualized knowledge of local dynamics. Here, I bring together diverse data sets to provide context to local characteristics ranging from soil data, to water bodies, access to mobile finance, and the quality of governance of local institutions, among many other variables. At the city-, neighborhood-, and household-level, sustainable decarbonization will not occur, or be equitable, without fully considering users, behavior, and co-designing technology (services and systems) that work for them and their communities. In this case, we use the nexus of information and communication technology, sensors, behavior and data mining to co-design information systems that simultaneously work for the user and the energy system in which they interact.
What are resource constrained environments? Resource constrained environments, here, are defined as spaces that exist in relative social, infrastructural, economic, or environmental scarcity, and can be found anywhere. Resource constrained environments can be found in California (e.g., Richmond in Contra Costa county) where low-income and predominantly African American residents are exposed to much higher concentrations of benzene, mercury and other hazardous pollutants due to the nearby Chevron refinery (12, 13), or Memphis, New Orleans and Birmingham where low- income households spend over 10% of their income on electricity (14). They can also be found in Kenya, where it may take a lot of time and money to reach rural communities to perform needs assessments and provision of basic services. Or, Managua (Nicaragua), where low, low-middle income neighborhoods can only afford used appliances, and have a panoply of barriers to access energy efficiency and sustainable energy services. Exploring, ideating, and implementing solutions for resource constrained environments requires knowledge of history, local context and dynamics, and many aspects of top-down (e.g., institutional, political) and bottom-up (e.g., end-users, neighborhoods) behavior.
This dissertation, explores demand-side, user-centered, sustainable decarbonization in resource constrained environments at multiple scales. The first chapter presents an analysis of global demand-side features that are enabling low-carbon transitions. It synthesizes 10 global energy and development related data sets and uses methods from data mining and concepts in behavioral science to propose a new methodology for the design and evaluation of long-term energy system decarbonization support mechanisms. Considering the nation-state as a single agent with its own historical intrinsic motivations for enacting change (e.g., social progress, environmentalism, economic efficiency, supremacy and empire), intrinsic characteristics (e.g., population size, land area, quality of governance), and enabling environments (e.g., local fuel and electricity prices, supporting policies) it uses these data to extract the features that can best explain decarbonization progress. We find higher local energy prices, foreign energy import dependency and absence of a large extractive resource base (e.g., oil and gas, mining), relative high investments in renewable energy (per km2 and capita), and early historical investments in geothermal energy and biomass for electricity to be key driving features. Policies, although widely advocated for in international frameworks, do not appear as key enabling drivers - especially in the rising south, and when misaligned with country specific motivators and intrinsic characteristics.
The second chapter explores and develops new methods in elucidating demand for the design and implementation of appropriate and sustainable energy interventions. It uses a mix of high spatial resolution data sets, surveys, sensor data, and data mining to develop new methods for elucidating rural electricity demand at the household and community level, and to help close the ‘energy efficiency gap’ in urban resource constrained environments. It uses an extended literature review and data at multiple scales to create a data-driven context that energy planners can use in their supply-side models. Using Kenya as an example, and with colleagues from the IBM-Africa (Nairobi) research lab, we develop what we consider to be the first reliable data-driven approach for elucidating household appliance ownership and induced household demand for electricity using a mixture of large-scale social demographic data, spatial data, and machine learning approaches. We also use data-mining and an extended literature review to explore and identify the enabling conditions under which electrification can lead to wealth via micro-enterprise creation in rural areas. The latter also presents the first analysis to evaluate the drawbacks/inaccuracies of the modern use of nightlights as a panacea for tracking wealth in unelectrified regions. Finally, and using Nicaragua as a case study, this chapter develops an extended literature review and framework on how to collect data for baseline energy efficiency estimates in resource constrained environments using a mixed methods approach combining surveys, sensors, population sampling and Bayesian updating.
The third chapter uses a field deployment pilot in Nicaragua as a case study, presenting opportunities and challenges for information and communication technologies (ICTs) and the internet of things (IOT) for demand-side flexibility and behavioral energy efficiency in resource constrained environments. We use ICTs and IOT to implement the first paired behavioral energy efficiency and flexible demand pilot in Latin America, in Nicaragua’s capital city of Managua. The chapter is divided in two sections, the first introduces the design, implementation, and exploratory data analysis of a sensor gateway (the FlexBox) for enabling behavioral energy efficiency and demand side flexibility, and the second is a post-implementation evaluation using Bayesian estimation for evaluating energy reduction, participation in demand side flexibility, impacts on welfare, and behavioral economics insights. We present several novel findings related to technology implementation, development of new efficiency parameters, and behavioral insights (e.g., incentive types, pre-existing behaviors, motivations) describing the opportunities and barriers to behavioral energy efficiency and demand side flexibility in these contexts. More importantly, we show that ICTs and IOT are mature technology that can be used by low, low-middle income households and small businesses in cities like Managua to become important actors in city-wide resource conservation. We conclude by presenting opportunities for future research.
