| dc.description.abstract | Sustainable operations has transformed in the past decade, as interest from consumers, companies and regulators has increased. There has been a growing excitement and necessity to leverage the large-scale data collected to improve the modelling and decision making around sustainable operations. In this thesis, we introduce new methodologies to support data-driven sustainable operations, and in specific deal with topics around electric vehicles and the COVID-19 pandemic.
In Chapter 2, we consider the problem of electric vehicles (EVs) as distributed storage for the electric grid. While electric vehicles can act as batteries supporting both the home and the electric grid, uncertainty around car usage must be accounted for before such models can be used in practice. We introduce a driver behavior-focused dynamic optimization for the charging and discharging of electric vehicles. We characterize policies that are interpretable to drivers to address distrust of automatized discharging of car batteries and prove analytically the regimes under which such policies are optimal. Finally, we work closely with an American EV manufacturer to show the dollar and carbon benefit that can be expected to be saved from discharging based on their driving behavior. We do this by clustering drivers based on their driving to derive probability distributions of when and how much drivers use their car to feed into the dynamic optimization.
We further develop the challenge of data-driven decision making in sustainability through Chapter 3. Rather than learning probability distributions as in Chapter 2, we introduce a deterministic approach in which a tree-ensemble model, specifically a random forest, forecasts how much drivers use their EV. This gives rise to a challenge from the predict-then-optimize literature around the tractability of optimizations in which an objective function is determined by a tree ensemble model. In this chapter we introduce an Upper Bounding Method for Optimizing over Tree Ensemble Models, UMOTEM. We demonstrate the scalability of UMOTEM, showing it grows linearly with regard to both the number of trees in the ensemble as well as those trees' depth. This is a strong improvement over comparable formulations which grow exponentially. We also bound the optimality gap introduced through the approximation, characterizing it using features of the random forest such as leaf separation and in-sample error. We computationally compare our approximation to similar methods, demonstrating that the algorithm captures over 90% of optimality in 2% of the runtime for publicly available datasets. Finally, we demonstrate the use of UMOTEM through two case studies. First, we take the same case as Chapter 2, and show how UMOTEM can be leveraged to optimize the charging and discharging of EVs. Second, we work closely with Oracle Retail to apply UMOTEM to promotion scheduling in order to determine an optimally markdown strategy for a fashion retailer.
In the final chapter of this thesis, we address data-driven decision making in one of the other major operational challenges to affect the globe, the COVID-19 pandemic. We develop a SIR-based model that can account for multiple waves. This model is agnostic to what drives the new waves (new variants, behavior changes, government policies, etc.) but takes a data-driven approach to identify when infection rates change. We prove analytical guarantees on how fast new waves can be detected. When modelling COVID-19, we show a new wave can be expected to be flagged within half a week. We also show strong computational results on COVID-19, demonstrating improvement over top COVID-19 forecasting models used by the CDC. | |
