Essays on Multi-Faceted Optimization and Data-Driven Modeling and Analysis of Logistics, Lot-Sizing, and Scheduling Problems

Abstract

This dissertation develops advanced optimization models and solution methodologies to address three complex and practically relevant problems in operations research: biomass feedstock logistics, integrated lot-sizing and scheduling, and data-driven large-scale production scheduling. Across these domains, the work contributes novel mathematical formulations, decomposition-based algorithms, and hybrid machine learning--optimization frameworks that improve the scalability, accuracy, and practical relevance of decision-making tools used in modern industrial and supply chain environments. The first part of the research proposes a comprehensive biomass logistics framework for cost-effective ethanol production from switchgrass. A mixed-integer programming model is developed that incorporates equipment routing, facility location, and transportation decisions, giving rise to a challenging three-stage structure. A nested Benders decomposition algorithm with multi-cut optimality conditions is introduced, enabling the solution of real-world problem instances that are otherwise intractable for state-of-the-art commercial solvers. Results demonstrate significant reductions in computational time and improved solution quality, thereby supporting the economic and environmental viability of biofuel supply chains. The second part investigates an uncapacitated integrated lot-sizing and scheduling problem with sequence-dependent setup costs. A network-flow-based formulation is proposed along with facet-defining inequalities that enhance model tightness. Deterministic, stochastic, capacitated, and backordering/lost-sales variants are developed, and a tailored Benders decomposition approach is introduced to solve large-scale instances efficiently. Computational studies show the proposed algorithm to outperform standard mixed-integer programming approaches and provide insights into the value of stochastic modeling for production planning under uncertainty. The third part addresses a real-life, large-scale integrated production scheduling problem in the manufacturing sector. A hybrid machine learning-mixed integer programming framework is developed to identify feasible machine-product assignments and generate optimal schedules that balance machine utilization, minimize changeovers, and satisfy operational constraints. Extensive tests using real production data demonstrate the approach's effectiveness and practical relevance. Together, these studies illustrate how rigorous mathematical modeling, advanced decomposition strategies, and data-driven insights can be integrated to solve high-dimensional optimization problems in logistics and manufacturing. The methodologies developed in this dissertation form a foundation for future research in stochastic optimization, machine learning-enabled decision systems, and large-scale industrial scheduling under uncertainty.