# Effects of physical properties estimation on process design a case study using AspenPlus.

код для вставкиСкачатьASIA-PACIFIC JOURNAL OF CHEMICAL ENGINEERING Asia-Pac. J. Chem. Eng. 2009; 4: 729–734 Published online 26 June 2009 in Wiley InterScience (www.interscience.wiley.com) DOI:10.1002/apj.328 Special Theme Review Effects of physical properties estimation on process design: a case study using AspenPlus Loic Cadoret,1 Cheng-Ching Yu,1 * Hsiao-Ping Huang1 and Ming-Jer Lee2 1 2 Department of Chemical Engineering, National Taiwan University, Taipei 106-17, Taiwan Department of Chemical Engineering, National Taiwan University of Sci. Tech., Taipei 106-07, Taiwan Received 31 October 2008; Revised 11 March 2009; Accepted 12 March 2009 ABSTRACT: This study focuses on the physical property model parameters estimation in order to accurately simulate separation processes for a given set of components. The non-random two-liquid (NRTL) model was chosen and parameters were calculated using different methods: experimental data regression and UNIFAC and COSMO-SAC (conductor-like screening model, segment activity coefficient) predictive models. The vapor-liquid equilibrium (VLE) obtained from these different models was compared and results showed that COSMO-SAC can be a reliable tool when data or functional groups are missing. Results also showed that the use of UNIFAC to estimate activity coefficients at infinite dilution can, in some cases, leads to inaccurate results and strongly impact process simulation results. 2009 Curtin University of Technology and John Wiley & Sons, Ltd. KEYWORDS: process simulation; thermodynamic properties; parameters estimation INTRODUCTION In the last decades, simulation software have turned out to be powerful tools to quickly design and improve chemical process. However, the lack or the wrong estimation of physical properties can easily lead to inaccurate results and non-negligible deviation from potential optimal process configurations.[1,2] The use of thermodynamic models can often provide good estimations of molecule interactions and the choice of the model appears to be of great importance. It will directly impact phase equilibrium calculations and thus separation units design.[3 – 6] Then, even with suitable thermodynamic models, the uncertainties in the model parameters can be significant. Through a literature review, Whiting[4] showed the consequences of inaccurate input parameters on process simulation results. According to the author, one should consider the entire system of data generation, model choice, parameters regression/estimation and simulation software choice to provide an actual accurate process optimization. Starting from a broad literature survey, Kister[7] identified the main sources of discrepancies between plant simulations and plant data. His study showed that for two-thirds of reported cases, main issues were related to poor predictions of vapor–liquid equilibrium (VLE) (especially for systems *Correspondence to: Cheng-Ching Yu, Department of Chemical Engineering, National Taiwan University, Taipei 106-17, Taiwan. E-mail: ccyu@ntu.edu.tw 2009 Curtin University of Technology and John Wiley & Sons, Ltd. exhibiting non-idealities and close-boiling molecules) and to the use of inconsistent experimental data. The author also recommended the use of graphical techniques (McCabe–Thiele and Hengstebeck diagrams) for simulations troubleshooting. Barnicki[8] showed that checking the quality of data was a critical step and that, consistency tests should always be performed to ensure accurate regressions and estimations of parameters. When using activity coefficient models [Wilson, UNIQUAC, non-random two-liquid (NRTL)], one often has to deal with a lack of experimental data and binary interaction parameters have to be estimated. The well-known UNIFAC group contribution method[9] is commonly used to solve this issue. In this model, molecules are subdivided into functional groups and thermodynamic properties are obtained by summing up contributions from all the pair-wise interaction groups. UNIFAC is a reliable method which gives good results, is extensively described in the literature, but has some drawbacks: experimental data are needed for fitting interaction group parameters and some functional groups remain missing. UNIFAC parameters are fit to a large number of data sets of components. By this averaging, for a given particular system, accuracy is not guaranteed and can lead to unnecessary over-design costs.[8] Among more recent methods to estimate physical properties, the conductor-like screening model-segment activity coefficient (COSMOSAC) method appeared to be very promising.[10,11] It 730 L. CADORET ET AL. Asia-Pacific Journal of Chemical Engineering is a new class of predictive model based on quantum chemical methods in which molecule surface is subdivided into segments with equal area and the chemical potential is derived from interactions between the surface segments of all molecules (a detailed description of the model can be found in Refs [11–13]). A significant advantage of this method is that no experimental data are needed to fit model parameters and thus, it can potentially handle any kind of system. In a previous study, Athès et al .[10] used NRTL and COSMO-SAC to predict VLE of aqueous mixture with aroma compounds. They showed that COSMO-SAC model was a reliable tool providing good estimations of VLE. Finally, for activity coefficient models, VLE predictions are very sensitive to activity coefficient (γ ) estimations, whatever the data sources (experimental data, estimation with UNIFAC or COSMO-SAC). The infinite dilution approximation is often used when activity coefficient values or not available for the whole range of composition. Barnicki[8] showed that an accurate estimation of γ at infinite dilution (γ ∞ ) for light-boiling component was a key for an accurate distillation simulation. The importance of γ ∞ in aqueous solution containing many aroma compounds was pointed out by Athès et al .[10] Authors showed that even for small amounts of aroma compounds, the total concentration of these molecules may not be negligible and γ may significantly deviate from γ ∞ , due to a synergy effect. Thus, the infinite dilution approximation has to be taken carefully. By comparing former cited models, this study aims at showing the importance of this estimation step in a process design. As a case study, a mixture of six main components was considered (Table 1). To illustrate impact of physical properties estimation, the separation of components by distillation was simulated using AspenPlus[14] software. This example was taken from a larger study conducted to design a whole new process.[15] It mainly dealt with the ammoximation of cyclohexanone, an intermediate step in the production of caprolactam. The chosen components are the main components involved in this process. ESTIMATION OF MODEL PARAMETERS The NRTL model[16] was chosen to represent both VLE and liquid–liquid equilibrium (LLE). This choice was justified by the presence of polar species, no electrolytes and a relatively low pressure (less than 2 bars) in the separation unit. The gas phase was considered as ideal with no vapor phase association. In the model (Eqns 1–3), binary interaction parameters (aij , bij and cij ) can be entered to accurately represent phase equilibrium (VLE and LLE) in the mixture: xj τji Gji j ln γi = + xk Gki k j τij − xj Gij xk Gkj k xm τmj Gmj m xk Gkj (1) k Gij = exp(−cij τij ) τij = aij + (2) bij T (3) With γi the activity coefficient of component i , xi the mole fraction, T the temperature and aij , bij , cij the binary parameters are to be estimated. These parameters are unsymmetrical, meaning that aij may not be equal to aji . For the cij parameter, it is assumed that cij = cji = 0.3 for all pairs of components, following the recommendations of Renon and Prausnitz.[17] It is worth noting that this is a simplified version of the model since Gij and τij could be expressed using six binary parameters.[14] A preliminary study showed that the three other parameters could be neglected. In some cases, the use of the six parameters even led to inaccurate results. In the literature, influence of the number of regressed parameters has already been observed.[4,18] It has been shown that number of parameters not only affects the VLE prediction but also, the simulation results (even without apparent discrepancies in VLE prediction). Table 1. Recapitulating available data. Boiling point NH3 −33.4 ◦ C t-butane Water Toluene c-hexanone Oxime PCES/CS O PCES/CS PCES/CS CS t-butanol 82.5 ◦ C Water 100 ◦ C Toluene 110 ◦ C c-hexanone 155 ◦ C Oxime 208 ◦ C O O Data CS O O CS O Data O – O, binary parameters available in the software; data, data regression; PCES and CS, binary parameters estimated with PCES software method and with Cosmo-Sac, respectively. 2009 Curtin University of Technology and John Wiley & Sons, Ltd. Asia-Pac. J. Chem. Eng. 2009; 4: 729–734 DOI: 10.1002/apj Asia-Pacific Journal of Chemical Engineering EFFECTS OF PHYSICAL PROPERTIES ESTIMATION ON PROCESS DESIGN Figure 1. Different ways used to obtain binary parameters for NRTL model in AspenPlus. Table 2. AAD measurements for different binary systems from the mixture. AAD System Regression PCES Cosmo-Sac Consistency test Oxime/toluene Oxime/cyclohxanone Water/t-butanol Water/ammonia t-butanol/cyclohexanone 0.012 0.193 0.0137 0.093 0.017 – – 0.045 0.205 0.13 0.011 0.291 0.046 0.13 0.035 Failed Failed Passed Failed Passed AspenPlus already provides parameter values for some pairs of components and, different ways were identified to obtain missing ones. The different procedures are summarized in Fig. 1. When available, experimental data were regressed. Thermodynamic consistency of data was first checked through area and point tests.[19,20] From VLE x -y curves, a mean deviation (AAD) between experimental and calculated points was evaluated, given by: AAD = N exp 1 |Yi − Yical | exp N i =1 Yi (4) exp Were Yi and Yical represent respectively experimental (retrieved from the literature sources) and calculated vapor molar fraction of component i , and N is the number of used experimental points. If VLE data are not available, binary parameters can be estimated based on activity coefficients for infinite dilution and constant temperature. For these conditions, NRTL model equations are simplified and the infinite dilution activity coefficient becomes only a function of the binary interaction parameters[14] : ln(γ1∞ , x1 → 0) = f (τ12 , τ21 ) (5) ln(γ2∞ , x2 → 0) = f (τ12 , τ21 ) (6) 2009 Curtin University of Technology and John Wiley & Sons, Ltd. Noting that for NRTL the third parameter cij is fixed, an easy system of two equations with two unknowns (aij , bij ) has to be solved. Obtained parameters are then used to calculate activity coefficients over the whole composition. Activity coefficients at infinite dilution can be given as experimental or estimated data using UNIFAC (method called PCES in AspenPlus). For this study, the more recent modified UNIFAC-Dortmund model[21] was used. As previously mentioned, UNIFAC cannot be used if some functional group is missing, and this is the case in this study for the cyclohexanone oxime molecule in which the ‘oxime group’ ( N-OH) is not available. In order to estimate interaction involving cyclohexanone oxime, the Cosmo-Sac model was used to calculate activity coefficients over the entire mole composition. Calculated coefficients are then used as ‘experimental data’ and regressed to calculate binary parameters. Depending on available data (Table 1), the different ways presented in Fig. 1 were tested and compared to each other via VLE calculation. AAD was used to evaluate deviation of estimation from experimental points. Results are presented in the following parts. Asia-Pac. J. Chem. Eng. 2009; 4: 729–734 DOI: 10.1002/apj 731 732 L. CADORET ET AL. RESULTS AND DISCUSSION Parameters regression from available experimental data Figure 2(a) and (b) shows results of regression for the two systems t-butanol/cyclohexanone and toluene/ oxime respectively. Mean deviations are recapitulated in Table 2. At first, it can be seen that experimental data did not pass thermodynamic consistency test for all cases. This means that a ‘blind’ trust cannot be put on retrieved data from literature. It then appeared useful to plot equilibrium curves to get more qualitative results. For the system toluene/oxime, despite the relatively low calculated AAD (1.2%), one can see the regressed bubble-point curve showing a significant deviation from experimental points. Since given data did not pass the consistency test, it becomes quite difficult to assess the accuracy of regression and show that even if data are available, efforts made to regress parameters do not necessarily lead to satisfying results. (a) (b) Figure 2. P-x-y VLE. (a) toluene/oxime pair (experimental data from Ref. [23]) (b) t-butanol/cyclohexanone pair (experimental data from Ref. [22]). 2009 Curtin University of Technology and John Wiley & Sons, Ltd. Asia-Pacific Journal of Chemical Engineering For the t-butanol/cyclohexanone system, the AAD is higher (1.7%), but calculated equilibrium curve follows well the experimental curve shape. These results suggest that a qualitative analysis of results is useful to assess for regression quality (as already mentioned in literature[22] ), especially when data are not consistent. AAD should thus only be seen as a semi-quantitative result, as it also strongly depends on the number of used experimental points. It is worth noting that in any case, compared to estimation methods, equilibrium retrieved from data regression are always closer to experimental points. Parameter estimation methods Figures 2 and 3 illustrate calculated VLE for different binary systems, using the different methods cited in previous part. These figures are representative of all results obtained with the different pairs listed in Table 1. As previously mentioned, experimental data are not available for all interacting pairs of components and UNIFAC could not be used for all pairs involving the cyclohexanone oxime molecule. In Fig. 2(a), (b) and 3(a), it can be seen that CosmoSac reproduces quite well experimental behaviors. For the t-butanol/water pair (Fig. 3(a)), the Cosmo-Sac model predicts well the azeotrope, with an AAD of 4.6%. When looking at curve shape, it even gives a relatively better result than UNIFAC, in a qualitative point of view. In these three cases, Cosmo-Sac remains very close to UNIFAC estimation. Figure 3(b) shows the result obtained for the ammonia/t-butanol pair. Here, the two models give quite different results. It can be observed that the use of UNIFAC with estimation of activity coefficients at infinite dilution (PCES) give quite peculiar results for small mole fraction of ammonia. This behavior was also observed using PCES for the ammonia/cyclohexanone pair. The reason for this result still do not appear clearly but shows that, this way of estimating parameters failed for these two systems, in these conditions. In literature, the sensitivity of results to activity coefficients at infinite dilution has already been observed. Athès et al .[10] showed that special care should be taken when employing the ‘infinite dilution’ approximation. The discrepancy observed on Fig. 3(b) could be related to this approximation when using UNIFAC to estimate γ ∞ (PCES). However in literature, discrepancies are usually observed for highly non-ideal system, close-boiling components[8] or when a synergy effect occurs.[10] In the presented case, the estimated VLE with COSMO-SAC do not exhibit a strong non-ideality and the two molecules do not have close boiling points. The two components are relatively ‘common’ and such result with UNIFAC is quite surprising. These results reinforced the fact that the use of Asia-Pac. J. Chem. Eng. 2009; 4: 729–734 DOI: 10.1002/apj Asia-Pacific Journal of Chemical Engineering EFFECTS OF PHYSICAL PROPERTIES ESTIMATION ON PROCESS DESIGN (a) (a) (b) (b) Figure 3. Vapor–liquid equilibrium. (a) t-butanol/water pair (experimental data from Ref. [22]) (b) ammonia/tbutanol pair. P = 0.5 atm. γ at infinite dilution has to be handled with care since it can lead to poor VLE estimations. As mentioned in the introduction, previous studies in the literature[10,11] showed that COSMO-SAC was able to provide accurate VLE estimations. Presented results are in agreement with former studies and show that COSMO-SAC can be more suitable than UNIFAC, for the specific mixture of interest. Thus, the model appears to be a useful alternative to the UNIFAC limitations (lack of functional group, inaccurate estimations). Impact on simulation Distillation of a four-component mixture was simulated with AspenPlus (Fig. 4). The separation takes place between water and the water/t-butanol azeotrope. Impact of the binary parameter estimation model was checked by keeping the same column configuration (reboiler duty and reflux ratio are kept constant) and changing the estimation method for one or more binary system. 2009 Curtin University of Technology and John Wiley & Sons, Ltd. Figure 4. Simulation results for the distillation of the ammonia/t-butanol/water/oxime mixture. For all component pairs, binary parameters were estimated following information given in Table 1, except for the ammonia/t-butanol pair, which was estimated via Cosmo-Sac (a) or PCES (b). Figure 4(a) and (b) shows results obtained by respectively using Cosmo-Sac and PCES for the ammonia/tbutanol system binary parameter estimation. With the use of Cosmo-Sac Fig. (4a), simulation predicted a clear separation with no ammonia and t-butanol in the bottom stream. The use of PCES Fig. (4b) led to a significant amount of these two light components in the bottom stream. This observed behavior with PCES is probably due to the peculiar VLE predicted for the ammonia/t-butanol system, as shown previously in Fig. 3. The strongly inaccurate estimation with PCES led to poor simulation results. It is worth noting that trusting PCES estimations would lead to a complete different approach for the Asia-Pac. J. Chem. Eng. 2009; 4: 729–734 DOI: 10.1002/apj 733 734 L. CADORET ET AL. Asia-Pacific Journal of Chemical Engineering process design and for the further separation of water and oxime. CONCLUSION Using AspenPlus simulation tool, the estimation of interaction parameters can be performed using different ways. At first, the regression of available data is not so obvious and has to be performed carefully. For missing parameters, UNIFAC and COSMO-SAC model were then compared. The COSMO-SAC model gave quite good results and appears to be of great interest when UNIFAC cannot be used. It was also shown that the estimation of activity coefficient at infinite dilution with UNIFAC could lead to inaccurate results and significantly, impact the process simulation. The influence of infinite dilution approximation on VLE predictions has already been mentioned in literature but still needs to be clarified in a future work. It can lead to poor estimations, even for components which do not exhibit close boiling point or strong non-ideality, as shown in this study. Finally, depending on the system of interest, every possible option should be considered to estimate parameters when data are not available, in order to avoid model failure and ensure process simulation accuracy. [2] S.G. Huang, C.C. Yu. J. Chin. Inst. Chem. Engrs., 2003; 34(3), 345–355. [3] A.R. Nelson, J.H. Olson, S.I. Sandler. Ind. Eng. Chem. Proc. Des. Dev., 1983; 22, 547. [4] W.B. Whiting. J. Chem. Eng. Data, 1996; 41, 935. [5] Y. Xin, W.B. Whiting. Ind. Eng. Chem. Res., 2000; 39, 2998. [6] P.M. Mathias, H.C. Klotz. Chem. Eng. Prog., 1994; 90(6), 67. [7] H.Z. Kister. Can We Believe Simulation Results? Chemical Engineering Progress, October 2002; 52–56. [8] S.D. Barnicki. How Good Are Your Data? Chemical Engineering Progress, June 2002; 58–67. [9] A.a Fredenslund, G. Gmehling, P. Rasmussen, Vapor-Liquid equilibria using UNIFAC, Amsterdam, Elsevier, 1977. [10] V. Athès, P. Paricaud, M. Ellaite, I. Souchon, W. Furst, Fluid Phase Equilib., 265, 139–154. [11] S.-T. Lin, S.I. Sandler. Ind. Eng. Chem. Res., 2002; 41, 899–913. [12] A. Klamt. J. Phys. Chem, 1995; 99, 2224–2235. [13] A. Klamt, F. Eckert. Fluid Phase Equilib., 2000; 172, 43–72. [14] Aspen Technology Inc.. ASPEN Plus User’s Manual, Aspen Technology Inc., Cambridge, 2000. [15] M. Kitamura, Y. Shimazu, M. Yako. Sumitomo Chemical Company, Limited. Continuous Process for Producing EpsilonCaprolactam by Gas-Phase Beckmann Rearrangement of Cyclohexanone Oxime, Patent Number EP 570,110, August 16, 2000. [16] J.M. Prausnitz, R.N. Lichtenthaler, E.G. de Azevedo. Molecular Thermodynamics of Fluid-Phase Equilibria 3rd edn, Taiwan, Pearson Education 2004. [17] H. Renon, J.M. Prausnitz. AIChE J., 1968; 14, 135–144. [18] A.K.S. Murthy, D. Zudkevitch. Inst. Chem. Eng. Symp. Ser., 1979; 56, 1.1-51–1.1-78. [19] O. Redlich, A.T. Kister. Ind. Eng. Chem., 1948; 40, 345. [20] M.M. Abbott, H.C. Van Ness. AIChE J., 1975; 21, 62. [21] U. Weidlich, J. Gmehling. Ind. Eng. Chem. Res., 1987; 26, 1372–1381. [22] J. Gmehling, U. Onken. Vapor-Liquid Equilibrium Data Collection, DECHEMA Chemistry Data Series. [23] K. Suehnel. Z. Phys. Chem., 1985; 9999. REFERENCES [1] E.C. Carlson. Don’t Gamble with Physical Properties for Simulations, Chemical engineering progress, October 1996. 2009 Curtin University of Technology and John Wiley & Sons, Ltd. Asia-Pac. J. Chem. Eng. 2009; 4: 729–734 DOI: 10.1002/apj

1/--страниц