(164c) Model-Based Filtering of Large-Scale Datasets - a Biorefinery Application Conference: AIChE Annual MeetingYear: 2014Proceeding: 2014 AIChE Annual MeetingGroup: Computing and Systems Technology DivisionSession: Advances in Intelligent Systems and Information Management Time: Monday, November 17, 2014 - 3:55pm-4:15pm Authors: Prunescu, R. M., Technical University of Denmark Sin, G., Technical University of Denmark Blanke, M., Technical University of Denmark Jakobsen, J. G., DONG Energy Numerous sensors monitor large scale plants and generate an enormous amount of data that are most often overlooked. This paper presents a methodology for processing raw data in real-time in order to obtain more useful process parameters. The method relies on building an inferential sensor based on a dynamic model of the system, and is applied on a demonstration scale biorefinery. The study includes results from processing real-time industrial data. In second generation biorefinery employing biochemical concepts, sugars are extracted from lignocellulosic agricultural wastes, which are then fermented in order to produce bioethanol. The biorefinery process consists of biomass pretreatment, fibers liquefaction or enzymatic hydrolysis, fermentation and separation steps . In 2013 the first commercial size biorefinery started to produce 2nd generation bioethanol  and several more sites are under construction nowadays, soon to be commissioned. Most biorefineries are monitored by trivial sensors, such as temperature, pressure, level, pH and flow measurements. The number of sensors in a typical commercial size plant can amount to several hundreds. It is impossible for process operators to monitor in real-time and make sense out of so many measurements. The operators normally extract historical values and post process them to calculate key parameters retrospectively, such as dry matter content in tanks, or mass and energy balances in different sections of the plant. In this way, sensors are fused together to produce fewer and better process indicators that help verify if the plant operates at desired optimal conditions or if the plant deviates from normal conditions, e.g. due to faults. Very often the raw historical data are filtered by trivial first or second order filters, and also downsampled in order to speed up calculations. Most filter parameters are found by cut and try design methods that eliminate noise. A more advanced technique of filtering and sensor fusing can be accomplished using a dynamic model of the plant and design observer-like feedback loops in order to accommodate real measurements into the model. An overview of such filter designs can be seen in . All these computations can be automatically performed fast enough in real-time nowadays, making real-time optimization and real-time fault detection possible. More than that, the dynamic model offers information that is not measured in reality, either because mounting specialized sensors in some points of the process can turn to be very expensive or because extracting samples from some reactors is nearly impossible, e.g. in steam pretreatment, the pressurized thermal reactor cannot be opened to extract samples and determine biomass composition. The dynamic model comprises ordinary differential equations (ODEs) that describe mass and energy balances for each individual unit from the process. If transportation and mixing effects become important, computational fluid dynamics techniques are employed for describing them properly , e.g. the convection diffusion reaction and the heat convection diffusion equations on one axis are used in the thermal reactor, which is a very long cylindrical shaped tank . Wherever steam exists, steam states are computed based on the IAPWS IF97 standard. Some units were also extended with fast pH and viscosity calculators, e.g. the liquefaction tank where enzymes are very sensitive to pH variations . To transform the model into an observer or inferential sensor, innovations or estimation errors are constructed for each measured signal. The innovations are then multiplied by a gain matrix and fed back into the model. The gain matrix can be designed either by pole placement or through Kalman procedures, i.e. basic, extended or unscented Kalman filters depending on system nonlinearities. The purpose is to reject process and measurement noise, and to offer an accurate estimate of the plant future state. This study uses basic Kalman filters where matrix gains are computed offline because an industrial large scale setup is expected to have constant operating points. The biorefinery inferential sensor is tested in real-time on a demonstration scale plant. Biomass composition in the pretreatment and liquefaction sections are estimated and compared to NIR and HPLC data, which were obtained offline in the laboratory. The estimates are in accordance with the laboratory analysis. This demonstrates that the observer is reliable for real-time optimization procedures or as soft measurements (as a soft sensor) for advanced control loops. References:  Larsen J, Haven M, Thirup L. Inbicon makes lignocellulosic ethanol a commercial reality. Biomass and Bioenergy, 2012; 46:36-45.  Beta Renewables opens biofuels plant in Italy. Focus on Catalysts, 2013; 12:6.  Khatibisepehr S, Huang B, Khare S. Design of inferential sensors in the process industry: A review of Bayesian methods. Journal of Process Control, 2013; 23:1575-1596.  Egeland O, Gravdahl JT. Modeling and simulation for automatic control. 2002.  Prunescu RM, Blanke M, Jensen JM, Sin G. Temperature Modelling of the Biomass Pretreatment Process. Proceedings of the 17th Nordic Process Control Workshop, 2012; 8-17.  Prunescu RM, Sin G. Dynamic modeling and validation of a lignocellulosic enzymatic hydrolysis process - A demonstration scale study. Bioresource Technology, 2013; 150:393-403.