Stochastic Programming Models to Integrate Biomass Quality Variability in the Design of Biofuel Logistics Networks

Biofuels are a promising alternative to replace the reliance on fossil fuels, due their sustainable production based on organic matter from living or recently living beings (i.e., biomass). The production of biofuels contributes to the reduction of greenhouse emissions (GHG), in particular CO2. According to government agencies such as the Energy Information Administration (EIA), an increment in biofuel production is expected in the U.S. within the next years. The 2017 Act of Congress encourages more production of fuels utilizing renewable resources such as biomass in the coming years.

Biofuels are classified according to their feedstock: (1) the first-generation is related to biofuels produced from biomass that can be generally used as food for human consumption, (2) the second-generation are biofuels generated from a wide range of feedstock, including lignocellulosic matter up to municipal solid waste and (3) the third-generation commonly refers to biofuels produced from algal biomass. Biofuels produced from lignocellulosic biomass (LCB) is a plausible alternative in the United States for the next years since there are not intended for human consumption and most of the production processes are already implemented by local producers.

In order to become a feasible alternative to their competitors (i.e., fuels from fossil-based production), biofuels require improvements in strategic areas such as conversion technologies, genetic manipulation of feedstock, supply chain of biomass from harvesting areas to conversion facilities, among others. Supply chain (SC) optimization constitutes an important opportunity area in biofuels economies since operations like harvesting, handling, storing, transportation, along with other processes have shown a significant impact on biofuel yields. Moreover, biomass properties play an important role in the design of the operations required for the production and distribution of biofuel.

Biomass has several properties that affect the current/future biomass conversion technologies. Moisture content is an example of a biomass physical property that impacts the conversion yield. In the case of the pyrolysis conversion, temperature is one of the main factors involved in the conversion, and thus, high content of moisture requires more energy to transform biomass in biofuel. Another example is the ash content effect in the thermochemical conversion, a high concentration of ash can impact the maintenance cost since the conversion process releases ash particles that accumulate in the boiler, and therefore, need to be removed to keep the equipment in operating conditions.

Biomass physical and chemical properties need to be considered in order to minimize the overall cost when design the biomass-to-biorefinery supply chain.

This dissertation proposes novel stochastic programming models to design such supply chain networks by introducing biomass quality variability, as well as other key operational factors, to improve the efficiency of biofuels logistics. This dissertation presents a compilation of three research papers (Chapters 1-3). In the first chapter, a hub-and-spoke model introduces variability in the biomass moisture and ash contents, to design a biofuel supply chain. A case study in the state of Texas is presented. The results show an impact on the investment and operation cost of approximately 8.31% due to the quality-related characteristics. Moreover, the optimal supply chain has a different configuration when biomass moisture content is considered. The case study is solved with a L-Shaped method with connectivity constraints, to accelerate the algorithm convergence. The problem is classified as NP-hard, and therefore, advanced algorithmic development is required to solve large-scale instances. The paper presented in Chapter 1 was submitted for publication to the journal Annals of Operations Research. This fact led to the second paper.

Chapter 2 extends the work in Chapter 1 by introducing an algorithmic novelty with the purpose of improving the solution time and the quality of the solution obtained previously. The methodology proposed in Chapter 2 combines metaheuristics and exact methods to create a solution procedure that utilizes a divide and conquer strategy to solve large-scale supply chain problems. The solution method has two main blocks. The first block provides an initial supply chain topology whereas the second block performs a fine-tune optimization of the hub-and-spoke network. The results significantly improve the run time and quality of the solution obtained with the L-shaped implementation (i.e., built-in CPLEX Benders Decomposition). The hybrid metaheuristic allows the decision maker to deal with higher levels of complexity regarding model formulations, and then, further features such as physical and chemical biomass properties can be included in the decision support process. Chapter 2 was submitted for publication to the journal of Cleaner Production.

Chapter 3 is a continuation of the work presented in Chapter 1 and Chapter 2 and it is a collaboration between the Mississippi State University (MSU) and the University of Texas at San Antonio (UTSA). Chapter number 3 introduces a comprehensive model with further considerations about biomass characteristics that affect the overall performance of the supply chain. The model contemplates seasonality in the biomass supply and links the dry matter loss with the time period between harvesting and preprocessing to get an accurate measure of the biomass degradation. Inventories at all levels are established to provide biomass supply during the year. Results from the extensive computational experimentation show that the incorporation of moisture, ash, and dry matter loss increases 44.44% the number of depots required in the network to densify the biomass and minimize the quality-related costs. Finally, an advanced enhanced progressing heading algorithm is proposed to solve a case study in the south-central region of the U.S., encompassing six states. The size of the problem reaches 2 million of variables (25,000 binary variables) and approximately 60,000 constraints, therefore, the problem is classified as a large-scale problem. Chapter 3 was submitted for publication to the journal of IISE Transactions.