Integration of machine learning and optimization for decision making under uncertainties with applications in agriculture and power system

Thumbnail Image
Akhavizadegan, Faezeh
Major Professor
Wang, Lizhi
McCalley, James
Hu, Guiping
Archontoulis, Sotirios
Olafsson, Sigurdar
Committee Member
Journal Title
Journal ISSN
Volume Title
Research Projects
Organizational Units
Journal Issue
Is Version Of
Industrial and Manufacturing Systems Engineering
This dissertation focuses on formulating and solving decision-making problems in uncertain environments using the integration of machine learning and optimization. Decision making under uncertainty is a challenging and complex problem because of logistical limitations. First, there exist enormous amount of uncertainty. Considering too many scenarios causes the model to be intractable, while too few scenarios cannot represent uncertainty enough. Second, due to time and resource constraints, it is unrealistic to design and validate algorithms in the real world. Hence, we need computer simulation. Third, considering multiple decision-making stages, many scenarios, complex interactions between scenarios, and decisions needs advanced algorithms and models. In an uncertain environment, the decision-making model has several real-world applications such as energy policies, agriculture, marketing, supply chain design, transportation, etc. The dissertation is devoted to both theoretical research on decision-making under uncertainty and its applications, which consists of three parts, each in a paper format. The first paper develops a new approach to select a small number of high-quality scenarios from many scenarios in the application of transmission expansion planning. Because of correlations between generation capacity, demand, and fuel cost, we develop a heuristic algorithm to capture the generation capacity and construct realistic scenarios. In this study, high-quality scenarios are chosen to minimize the Kantorovich distance of social welfare distributions between the selected and the whole set of scenarios. We explore the performance of the proposed framework on U.S. Eastern and Western Interconnections as a case study. In the second paper, we focus on parameter calibration of computer simulation to have a realistic model for designing and validating the proposed decision-making framework. Because calibration of simulation software is required to reflect nature more realistically, we present a new automated framework and a parallel Bayesian optimization algorithm to estimate time-dependent parameters. In this paper, we focus on a crop model as the simulator of nature. A new automated framework by integrating the power of optimization and machine learning with agronomic insight is proposed to tune time-dependent parameters for crop models and have a realistic simulator. The third paper develops risk-averse stochastic optimization frameworks to optimize management practices and select the best cultivar at different levels of risk aversion. We integrate the crop model and an optimization algorithm to develop multiple stages of the decision-making process. The optimization framework at different levels of risk aversion contains an optimizer and a simulator. The optimizer uses parallel Bayesian algorithms as the core search engine to effectively search the best combinations of management decisions over unknown objective functions. The crop model as a simulator is used to capture complex interactions between scenarios and decisions and evaluate optimizers’ decisions appropriately. The objective function in this study is maximizing yield by optimizing planting date, N fertilizer amount, fertilizing date, and plant density in the farm, and selecting the best cultivar with different maturity days. As a case study, we use a crop model of 25 locations with different environments across the US Corn Belt and optimize for three test years (2016-2018) at three time-wise strategies during the growing season.