Robust Prediction of Gasoline Properties Using Spectroscopic Methods Dynamically Corrected with Lab Data, A Barsamian

Tags: Ara Barsamian, spectroscopic method, NIR analyzer, Lab data, properties, property, practical process control, parameter determination, model parameter, Lab Data Ara Barsamian Refinery Automation Institute, LLC, chemometric, gasoline octane, property prediction, model prediction, predicted properties, blend component, flow volumes, component properties, spectroscopic measurement, Barsamian, volumetric calculations, Spectroscopic Methods
Content: Robust Prediction of Gasoline Properties Using Spectroscopic Methods Dynamically Corrected with Lab Data Ara Barsamian Refinery Automation Institute, LLC (973)-644-2270 [email protected] 1. Introduction Spectroscopic methods for measuring in-line blended gasoline have many advantages over conventional on-line analyzers, such as very fast, very precise measurements, simultaneous multiple parameter measurements, and very high reliability, at a reasonable cost. Spectroscopic analysis methods in this paper refer only to FTIR, Raman, and magnetic resonance Analysis. The major shortcomings of using spectroscopic methods or any inferential model-based parameter determination are: inaccurate inferential model predictions of parameters of interest, e.g. gasoline octane values, lack of a practical, how-to guide to build simple but reasonable inferential (chemometric) models. Lack of a simple methodology to dynamically correct the predicted parameter value of an imperfect inferential model by comparing it with a credible value, e.g. Lab This paper addresses these problems with a robust solution and scheme for reliable gasoline (or any fuel) blenD Property parameter prediction (chemometric) model development method, and dynamic correction of the model parameter prediction with Lab data. The scheme is based on well-known principles of inferential model-based prediction of process stream quality in place of a real parameter measuring device, e.g. octane knock engine. The model output is corrected periodically with Lab data to within the ASTM reproducibility of that parameter. The described approach has been used since 1960's with the advent of practical Process Control computers to implement composition control without using on-line property analyzers [1], and later to validate octane knock engines and NIR-type spectroscopic analyzers since 1986 [2, 3], and is a derivative of tank quality integration used in in-line blend property control since its inception in 1965 [4]. The scheme is valid for: Any fuel blending (gasoline, diesel, bunker), or any mixture of liquids Any parameter for which an inferential model can be developed, e.g. AKI, RVP, etc. In-line fuel blending scheme, either rundown blending or component tank blending Notice: The information, methods and techniques described in this paper is OPEN, Public domain information. It cannot, and should not be patented, or in any way oblige the public to pay any royalty fees to any user or implementer of the described methods and techniques. ©2016 by Ara Barsamian. All rights reserved.
2. Summary of the "Robust" Scheme The elements of the scheme are shown in Fig. 1, consisting of a inferential model used to predict process stream quality in place of a conventional on-line analyzer, e.g. octane knock engine, but corrected dynamically in real-time with Lab data . The inferential model, in our case a spectroscopic chemometric model, requires inputs of process data such as FTIR transmission or absorbtion spectrum, Raman spectrum, or magnetic resonance, etc. The model predicted parameter is periodically updated by comparing the predicted value(s) with Lab value(s). It is important to set limits to the applicability of the model to avoid gross errors, in our case, the model output is bounded by the ASTM precision parameter (r and R) Fig. 1 Principle of spectroscopic method property prediction correction scheme The implementation of the scheme comprises the following four steps: Step 1: Develop a multivariate property prediction model using conventional commercial chemometric software. The inputs for model building are the monthly gasoline recipes derived from a refinery-LP annual business plan; the recipes are hand-blended in the Lab to establish recipe vs. spectrum correlation. This model does not need to be of very high fidelity, since its predictions will be corrected in real-time by Lab data, and bound by ASTM r&R to avoid gross errors. Notice: The information, methods and techniques described in this paper is OPEN, Public Domain information. It cannot, and should not be patented, or in any way oblige the public to pay any royalty fees to any user or implementer of the described methods and techniques. ©2016 by Ara Barsamian. All rights reserved.
Step 2: Use in-line blending data reconciliation algorithm whereby we synthesize the in-line blended gasoline properties from blend component refinery Lab LIMS computer stored property data multiplied by Flow volumes measured by blender flowmeters for each component (similar to quality integration with FPAPV described in ASTM D6624), and summing the result. The result is a calculated value of the properties of the blended liquid at the in-line blend header based on blend component Lab property data and flow volumes measured by flowmeters. This is done at the same frequency as the spectroscopic measurement ( every 1 to 2 minutes) Step 3: We then compare the Lab-based "synthetically" calculated properties of the blend liquid, (e.g. gasoline) in the blend header from Step 2 against the spectroscopic method model prediction in Step 1. This comparison is done at the same frequency as the spectroscopic measurement (every 1 to 2 minutes) Step 4: We take the results from Step 3 and we add a correction to the spectroscopic chemometric prediction while insuring that we set limits to the applicability of the model (i.e. within ASTM "r and R") to avoid gross errors, as follows: Case 1 - If the difference between the value of the spectroscopic model prediction and "synthetic" volume average calculation is less than the ASTM repeatability, r, the Step 1 prediction is valid and can be used for property control or certifying a blend. Case 2 - If the difference between the value of the spectroscopic model prediction and "synthetic" volume average calculation is greater than than the ASTM repeatability, r, but less than the ASTM Reproducibility of the parameter, then we correct the Step 1 prediction by adding the difference between Step 1 and Step 2, as long as the difference is less than the ASTM Reproducibility of the parameter. Case 3 - If the difference between the value of the spectroscopic model prediction and "synthetic" volume average calculation is greater than the ASTM Reproducibility of the parameter, R, then we cannot use the Step 1 prediction, which is invalid and cannot be used for blend property control or certifying a blend. This requires a "root-cause" analysis to determine the cause, e.g. an outlier due to an unusual blend component or unusual recipe. 2. Spectroscopic Method Blend Property Prediction Model Development The spectroscopic chemometric prediction model building "how-to Bible" it the ASTM E1655 practice. Contrary to claims in literature, the challenge with it is its practical implementation to cover the "representative recipe space" for the fuel grade of interest, which is beyond the "ordinary mortal" in a refinery, commercial "for-hire" blending facility, or spectroscopic analyzer/device supplier. Why is that? Because blend recipes, with all the good will and effort, are not consistent and predictable even for the same grade and season, depending on process unit sometimes lousy operation causing significant property variability, components on hand, occasional opportunistic cargos with poorly know properties, inventory and storage constraints forcing unscheduled blends, and poorly configured planning tools (both at refinery LP level and blend planning tools, like Multi and Single Blend Optimizers, nonlinear blending, Ethanol boost, etc.). Notice: The information, methods and techniques described in this paper is OPEN, Public Domain information. It cannot, and should not be patented, or in any way oblige the public to pay any royalty fees to any user or implementer of the described methods and techniques. ©2016 by Ara Barsamian. All rights reserved.
To circumvent these headaches, we pursue a different approach: build an approximate (lousy?) model which is periodically corrected with Lab data, with the model output bounded by the ASTM r&R for each predicted parameter. Development of a multivariate property prediction model uses well-recognized, high quality commercial chemometric software, as follows: 2.1 The inputs for model building The inputs for model building are the monthly gasoline recipes derived from a refinery-LP annual business plan; the recipes are hand-blended in the Lab to establish recipe vs. spectrum correlation. The recipes are per grade, but cover all the seasons (Summer, Transition, Winter). The example in Fig. 2 indicates potential for further simplification, since examination of recipes for the Winter months shows four almost identical recipes which we can reduce them to one, and the same considerations applies to Summer gasolines. Fig. 2 Refinery LP Annual Business Plan Gasoline Monthly Average Recipes for CG87 for 2004 If you contemplate blending RBOB's and CBOB's, you have two additional choices: 1) may wish to add 10 v% Ethanol to the "neat" hand blends to build also a model for direct Ethanol Gasoline "match" blends 2) regress the 10 v% Ethanol hand-blends to come up with "Ethanol Boost" prediction equations that can be pre-programmed in the spectroscopic analyzer measuring the neat blend, and then using them to calculate the resulting "boost" in octane, RVP, etc. 2.2 The Lab Hand Blends for Model Building This step requires a Lab to hand blend the recipes derived from step 2.1 The most challenging part is blending Butane (LPG!) in the hand blend flask, which is a lost art... Notice: The information, methods and techniques described in this paper is OPEN, Public Domain information. It cannot, and should not be patented, or in any way oblige the public to pay any royalty fees to any user or implementer of the described methods and techniques. ©2016 by Ara Barsamian. All rights reserved.
Once you have the hand blend, you do two things with the hand blend sample: 1. Conventional Lab analysis of the hand blend properties 2. Obtaining the spectra of the hand blend sample by using a Lab-installed spectroscopic analyzer These are part of the input to the spectroscopic chemometric model-building software 3. Development of FPAPV Property (Quality-Barrels) integration algorithm In this step, we use in-line blending data to calculate a prediction of the in-line blended gasoline properties from blend component refinery Lab LIMS-stored property data, Q, multiplied by flow volumes measured by flowmeters, F, for each component to get the property-barrels (similar to classical Tank Quality Integration (TQI), or alternatively ASTM D6624 quality integration method called FPAPV, Flow-Proportioned Average property value for a Collected Batch of Process Stream Material Using Stream Analyzer Data ), and summing the result. Fig. 3 Scheme for doing the TQI integration calculations for FPAPV The prediction could be improved somewhat by taking into account analyzer dead time and transport lag into account, using run of the mill dead-time compensators. 4. Putting It All Together After doing steps 2) and 3) above, we have all the pieces we need to do the confidently the actual blend property prediction. The display (Fig. 4) summarizes the actual spectroscopic analyzer performance, using the synthetic prediction of what is in the in-line blender header, the spectroscopic chemometric model prediction (based on the recipes), the comparison calculation, and corrections, if need be. Notice: The information, methods and techniques described in this paper is OPEN, Public Domain information. It cannot, and should not be patented, or in any way oblige the public to pay any royalty fees to any user or implementer of the described methods and techniques. ©2016 by Ara Barsamian. All rights reserved.
The overall scheme relies on comparing the "synthetically" predicted properties of the gasoline in the blend header from Step 3 against the spectroscopic method model prediction in Step 2. Fig. 4 Spectroscopic (NIR) Analyzer performance monitoring Display The real-time analyzer monitoring display labels are: ANALYZER= raw readings from the actual NIR analyzer chemometric model; these might be analyzer lag and transport lag corrected PREDICTION=predicted blend header properties synthesized using FPAPV "octane-barrels" volumetric calculations using flowmeter and blend component properties measured in the Lab BIAS=difference between NIR analyzer readings and prediction calculations CORRECTED=NIR analyzer readings corrected by biasing results depending on difference between ANALYZER and PREDICTION magnitude using ASTM "r or R" as described above To summarize, there are 3 comparison cases: 1. If the difference between the value of the spectroscopic model prediction and "synthetic" volume average calculation is less than the ASTM repeatability, r, the Step 1 prediction is valid and can be used for property control or certifying a blend. Notice: The information, methods and techniques described in this paper is OPEN, Public Domain information. It cannot, and should not be patented, or in any way oblige the public to pay any royalty fees to any user or implementer of the described methods and techniques. ©2016 by Ara Barsamian. All rights reserved.
2. If the difference between the value of the spectroscopic model prediction and "synthetic" volume average calculation is greater than the ASTM repeatability, r, but less than the ASTM Reproducibility of the parameter, then we correct the Step 1 prediction by adding the difference between Step 1 and Step 2, as long as the difference is less than the ASTM Reproducibility of the parameter. 3. If the difference between the value of the spectroscopic model prediction and "synthetic" volume average calculation is greater than the ASTM Reproducibility of the parameter, R, then we need to do a "root cause analysis" to determine the source of error. We cannot use the Step 1 prediction, which is invalid and cannot be used for blend property control or certifying a blend. 5. Conclusion The method described above is robust because it depends primarily on the Lab-measured properties of the blend components and reliable flowmeters, and only secondarily on the spectroscopic chemometric model. It was successfully used in an in-line gasoline blending project since 1998. The system used a NIR analyzer measuring 10 properties once a minute. The results were compared against FPAPV using turbine meters and blend component Lab data; the refinery Lab verified the long term precision and stability of the spectroscopically determined parameter prediction. REFERENCES [1] Wherry, T.C, and Parsons, J. R., "Guide to Profitable Computer Control", Hydrocarbon Processing, Vol. 45, No. 4, April 1967, pp. 71 [2 Diaz, A, and Barsamian, J. A., "Meet Changing Fuel Requirements with Online Blend Optimization", Hydrocarbon Processing, Vol. 75, No. 2, February 1996, pp. 179 [3] Popkowski, A., Barsamian, A., "New gasoline-blending unit started at Poland refining", Oil & Gas Journal, V.97, No. 27, July 5, 1999 [4] Weiland, R., "Computer Control of Motor Gasoline Blending", Paper 690226, presented at Society of Automotive Engineers, Jan. 13, 1969. Similar paper presented by R.J. Lasher at NPRA Computer Conference in 1965! Notice: The information, methods and techniques described in this paper is OPEN, Public Domain information. It cannot, and should not be patented, or in any way oblige the public to pay any royalty fees to any user or implementer of the described methods and techniques. ©2016 by Ara Barsamian. All rights reserved.

A Barsamian

File: robust-prediction-of-gasoline-properties-using-spectroscopic.pdf
Title: Spectroscopic models for certifying octanes
Author: A Barsamian
Author: Ara Barsamian
Subject: Chemometric models for Gasoline NIR
Published: Sun Jun 26 14:08:22 2016
Pages: 7
File size: 0.25 Mb


OCB Mode, 22 pages, 0.14 Mb
Copyright © 2018 doc.uments.com