Dynamic Modeling, Optimization, and Advanced Control for Large Scale Biorefineries

Abstract

Second generation biorefineries transform agricultural wastes into biochemicals with higher added value, e.g. bioethanol, which is thought to become a primary component in liquid fuels [1]. Extensive endeavors have been conducted to make the production process feasible on a large scale, and recently several commercial size biorefineries became operational: Beta Renewables (Italy, 2014), Abengoa Bioenergy (USA, 2014), POET-DSM (USA, 2014), GranBio (Brazil, 2014) [2], while others are under construction, e.g. the Måbjerg Energy Consortium in Denmark. This thesis presents the findings of a 3 years PhD project that was run by Technical University of Denmark (DTU) in collaboration with the largest Danish energy company DONG Energy A/S between 2012 and 2015. The company owns a demonstration scale second generation biorefinery in Kalundborg, Denmark, also known as the Inbicon demonstration plant [3]. The goal of the project is to utilize realtime data extracted from the large scale facility to formulate and validate first principle dynamic models of the plant. These models are then further exploited to derive model-based tools for process optimization, advanced control and real-time monitoring. The Inbicon biorefinery converts wheat straw into bioethanol utilizing steam, enzymes, and genetically modified yeast. The biomass is first pretreated in a steam pressurized and continuous thermal reactor where lignin is relocated, and hemicellulose partially hydrolyzed such that cellulose becomes more accessible to enzymes. The biorefinery is integrated with a nearby power plant following the Integrated Biomass Utilization System (IBUS) principle for reducing steam costs [4]. During the pretreatment, by-products are also created such as organic acids, furfural, and pseudo-lignin, which act as inhibitors in downstream processes. The pretreated fibers consist of cellulose and xylan, which are then liquefied in the enzymatic hydrolysis process with the help of enzymes. High glucose and xylose yields are thus obtained for co-fermentation. Ethanol is recovered in distillation columns followed by molecular sieves for achieving a high concentration ethanol. Lignin is separated in the first column and recovered as bio-pellets in an evaporation unit. The bio-pellets are then burnt in the nearby power plant for steam generation. The first part of this research presents a large scale dynamic model of the plant, separated in modules for pretreatment, enzymatic hydrolysis, and fermentation. The pretreatment and enzymatic hydrolysis models have been validated and analyzed in this study together with a comprehensive sensitivity and uncertainty analysis [5, 6]. The models embed mass and energy balances with a complex conversion route. Computational fluid dynamics is used to model transport phenomena in large reactors capturing tank profiles, and delays due to plug flows. This work publishes for the first time demonstration scale real data for validation showing that the model library is suitable for optimization, control and monitoring purposes. As an application, the pretreatment dynamic model is used to construct a realtime observer that acts both as a measurement filter, and soft sensor for biomass components that are not measured, e.g. pretreatment inhibitors [5]. The next part of this study deals with building a plantwide model-based optimization layer, which searches for optimal values regarding the pretreatment temperature, enzyme dosage in liquefaction, and yeast seed in fermentation such that profit is maximized [7]. When biomass is pretreated, by-products are also created that affect the downstream processes acting as inhibitors in enzymatic hydrolysis and fermentation. Therefore, the biorefinery is treated in an integrated manner capturing the trade-offs between the conversion steps. Sensitivity and uncertainty analysis is also performed in order to identify the modeling bottlenecks and which feedstock components need to be determined for an accurate prediction. This analysis is achieved with Monte Carlo simulations and Latin Hypercube Sampling (LHS) on feedstock composition and kinetic parameters following the methodology from [5, 6, 8, 9]. In the last part of this work, two applications of the L1 adaptive output feedback controller [10] are developed: one for biomass pretreatment temperature [11] and another one for pH in enzymatic hydrolysis [12]. Biomass conversion is highly sensitive to these process parameters, which exhibit nonlinear behavior and can change nominal values. The adaptive controllers are found to perform better across multiple operational points without the need of retuning.