Challenges and Prospects for Data-Driven Climate Change Mitigation

2019-03-04T20:11:24Z (GMT) by Lynn Kaack
Successful climate change mitigation will require data-driven decision making, but the field faces a diverse set of challenges. In this dissertation, I provide three examples that illustrate how uncertainty is often not adequately characterized, how missing data can pose a barrier to climate-relevant policy making, and how big data and machine learning could be used to obtain important information. I conclude with a survey and a discussion of how artificial intelligence can be applied to climate change mitigation. In the first chapter, I show how to construct an empirical estimate of the uncertainty of long-term<br>energy forecasts based on past forecast errors, using projections made by the U.S. Energy Information<br>Agency (EIA). This method gives analysts and decision-makers a means to estimate the uncertainty<br>of those forecasts quantitatively. Energy forecasts provide the basis for financial evaluation of energy<br>investments as well as for energy system models. I lay the groundwork for evaluating the performance of these methods in the data-scarce setting of long-term forecasts. The EIA has used my results in their most recent retrospective review. The second chapter is based on a topical review of policies to decarbonize heavy freight transportation<br>by shifting freight from road to rail and water. I find that while the freight sector is responsible for a large share of greenhouse gas (GHG) emissions, a systematic analysis of the potential to decarbonize with modal shift is still missing from the literature. This is partly due to a lack of<br>publicly available, standardized, and updated data. For a global comparison of modal split and trends, I expanded existing databases with national freight activity from 2000-2017. I find that only less than half of the countries in the world provide such information on road freight activity. The third chapter provides an example of how big data and machine learning (ML) could be used to fill in information gaps that inhibit climate policy analysis. I use satellite imagery for truck traffic monitoring in areas where this information is otherwise difficult to obtain. I count the number of freight vehicles visible in the images with deep convolutional neural networks, and estimate the<br>average annual truck traffic on roads from those counts by modeling traffic variation patterns. In a final chapter, I discuss how methods from artificial intelligence can be used to improve socioeconomic, policy, and engineering research for climate change mitigation. I provide a survey of<br>the literature and identify the main barriers and challenges that arise at the intersection of those disciplines. Research in this area demands both careful design of ML algorithms and consideration of domain knowledge. I conclude with proposing a research agenda. <br>