- Email: [email protected]
Reid vapor pressure prediction of automotive gasoline using distillation curves and multivariate calibration
Fuel 187 (2017) 167–172
Contents lists available at ScienceDirect
Fuel journal homepage: www.elsevier.com/locate/fuel
Full Length Article
Reid vapor pressure prediction of automotive gasoline using distillation curves and multivariate calibration Gisele Mendes a, Helga G. Aleme b, Paulo J.S. Barbeira a,⇑ a Laboratório de Ensaios de Combustíveis, Departamento de Química, ICEx, Universidade Federal de Minas Gerais, Av. Antônio Carlos 6627, 31270-901 Belo Horizonte, Minas Gerais, Brazil b Departamento de Ciências Exatas e da Terra, ICAQF, Universidade Federal de São Paulo, R. São Nicolau 210, 09913-030 Diadema, São Paulo, Brazil
a r t i c l e
i n f o
Article history: Received 15 June 2016 Received in revised form 26 August 2016 Accepted 12 September 2016
Keywords: Vapor pressure Gasoline PLS Distillation
a b s t r a c t Partial least squares regression (PLS) in conjunction with distillation curves (ASTM D86) were used successfully to predict the vapor pressure of automotive gasoline. The errors obtained in the calibration and validation models (0.71 and 0.69 kPa) were lower than those reported in the literature. In addition, the proposed method is inexpensive, reduces test time and is easy to implement, making it an alternative method for gasoline quality control. It does be possible because the distillation tests are already routinely performed by ANP as one of the evaluation parameter of the automotive gasoline quality. Ó 2016 Elsevier Ltd. All rights reserved.
1. Introduction Automotive gasoline is a flammable liquid formulated by mixing hydrocarbons ranging from four to twelve carbon atoms, with boiling points between 30 and 220 °C, obtained by refining petroleum, involving direct distillation, cracking, reforming, isomerization and alkylation [1,2]. Its composition varies according to the method employed in its production and the chemical composition of the crude oil. In 2015 about 44.2 billion liters was sold in Brazil [3]. In this context, there is an increasing need for quality control of gasoline and the development of new technologies to ensure that the product meets the requirements of engines and does not produce emissions above legal levels [4]. According to the ANP (National Petroleum, Natural Gas and Biofuels Agency), various parameters are used in quality control of gasoline, such as distillation parameters, specific gravity, and octane numbers, among others [5]. Currently, the analyses of these parameters are carried out in Brazil using the ASTM (American Society for Testing and Materials) or NBR (Brazilian Association of Technical Standards) standard methods. A fuel must meet all the specifications to be considered of good quality. An important physical-chemical parameter in assessing the quality of gasoline is the vapor pressure, because it is related to the fuel’s volatility, which directly influences engine performance. ⇑ Corresponding author. E-mail address: [email protected] (P.J.S. Barbeira). http://dx.doi.org/10.1016/j.fuel.2016.09.046 0016-2361/Ó 2016 Elsevier Ltd. All rights reserved.
In the production of fuel, vapor pressure is measured to indicate the requirements that must be met for the transport and storage of the product, in order to avoid accidents and minimize losses by evaporation. Fuels with high vapor pressure emit more volatile compounds and, after production and shipping fuel to service stations, this test can be applied to evaluate nonconformities related to adulteration caused by adding solvents. Takeshita et al. [4] noted that the addition of organic solvents at different ratios modifies the vapor pressure of gasoline, an effect that is more preponderant at 50% (v/v) and above. The evaluation of vapor pressure is also required to assure proper engine performance, since fuels with low vapor pressures are harder to vaporize in the intake manifold, hindering the combustion process. On the other hand, fuels with high vapor pressure emit more volatile organic compounds, which can cause incomplete fuel vaporization and uncontrolled air/fuel ratio. Hence, the vapor pressure must be within an acceptable range [6]. In Brazil, ANP establishes a maximum of 69.0 kPa for the vapor pressure, at 37.8 °C, of automotive gasoline [5]. The volatility parameters, which affect the performance at low temperature, are typically determined by testing vehicles at high RPM. At high temperatures, the most common phenomena that affect fuel systems in vehicles are vapor lock and percolation. Vapor lock is the occurrence of a mass of vapor between the fuel tank and carburetor or fuel injection system. Percolation is the result of uncontrolled fuel vaporization, which can happen after a long period operating at high temperatures [6].
168
G. Mendes et al. / Fuel 187 (2017) 167–172
Vapor pressure can be measured following ASTM D5191 [7]. The test conditions are carefully specified since the vapor pressure varies with temperature, the amount of dissolved air and the ratio between vapor and liquid in the container. In this method, the sample contained in a jar (filled to between 70 and 80% of capacity) is cooled to temperatures between 0 and 1 °C and subsequently inserted into the liquid chamber and the equipment is connected to the vapor chamber heated to 37.8 °C. The increased pressure in the vapor chamber is measured using a sensor transducer and a pressure indicator and the vapor pressure is registered on the display. This measure is defined as Reid vapor pressure (RVP) and can be expressed in kPa or psi. Although the test based on ASTM D5191 be relatively quick (approximately seven minutes), it is desirable to have simpler and low cost methods for RVP determination. In this context, chemometric methods combined with distillation curves can provide important information about different gasoline quality parameters, reducing the number of tests to being performed and used as screening for more specific tests. The current literature demonstrates the predictive power of multivariate regression models for vapor pressure prediction of automotive gasoline [8–11]. Among the regression techniques available, the most widely used are partial least squares regression (PLS) and principal components regression (PCR). Cooper et al. [8] correlated the results of vapor pressure obtained experimentally together with the Raman spectra of fuels to build a regression model using PLS. The authors found RMSEP values (root mean squares error of prediction) of 3.92 kPa. Using dispersive fiber-optic Raman spectroscopy with a CCD detector and near infrared, Flecher et al. [9] determined the vapor pressure of gasoline samples, finding RMSEP of 5.99 kPa. Besides PLS, Côcco et al. [10] associated near-infrared spectroscopy and the chemometric tool PCR to determine different chemical and physical parameters, including vapor pressure. For this property, the value of RMSEP obtained was 2.81 kPa. Recently, several works have demonstrated the good potential of distillation curves in conjunction with PLS regression to forecast quality parameters, such as specific gravity and ethanol content [11], MON and RON in automotive gasoline [12], flash point [13], cetane number [13], biodiesel content [14], specific gravity [15] and kinematic viscosity [15] in diesel oil. In addition to these parameters, these models have the ability to discriminate the refinery of origin [16] and to detect adulteration by solvents [17]. Distillation is a method of separation based on the phenomenon of vapor-liquid equilibrium for mixtures. With the ASTM D86 method it is possible to evaluate the complexity of liquid blends, which is directly related to the volatility of the sample’s components [18]. To simplify the analysis of gasoline through the use of a single test, thus reducing costs and increasing the number of samples that can be analyzed, this work describes the use of distillation curves from regular tests to evaluate the quality of gasoline, obtained according to ASTM D86 [18], in conjunction with multivariate calibration PLS for determination of the vapor pressure. We carried out manual distillations to obtain the composition of each fraction using infrared spectrometry with the purpose of showing the most important variables in predicting vapor pressure.
2. Experimental 2.1. Samples To build the PLS model used in the determination of the vapor pressure of gasoline, 80 samples of premium and regular gasolines
(both containing ethanol 25 % v/v) were obtained from service stations in the eastern region of the state of Minas Gerais (Brazil), which receives fuels produced by five different refineries. A study proposed by Aleme et al. [16] showed that gasolines produced by these refineries have different distillation curve profiles, due to their different chemical composition, related to the source of crude oil and the refining process thereof. The samples were stored in polyethylene bottles, sealed and chilled to 8–15 °C to avoid evaporation of volatile components. All samples were subjected to the tests of physico-chemical parameters established by the ANP [5]. 2.2. Materials and equipment 2.2.1. Vapor pressure analysis Before starting the tests, the samples were transferred to 50 mL flasks, filled to between 70 and 80% capacity and sealed. The samples were cooled in an ice-water bath to a temperature between 5 and 7 °C. After reaching the proper temperature, the flasks were shaken and opened momentarily. This process was repeated three times. Before the readings, the measuring chamber was cleaned with toluene and dried with n-pentane with the aid of a vacuum pump. After ensuring that the measuring chamber was completely dry, one of the valves was closed and a volume of 2.55 mL of air was introduced with a syringe, after which the valve was closed quickly to achieve pressure between 19.0 and 21.0 kPa. After one minute for stabilization, 2.05 mL of sample was and injected into the measuring chamber, with temperature of 37.8 ± 0.1 °C, using a syringe. The results were displayed in approximately three minutes, in kPa. The test was repeated three times for each sample. The reagent npentane (known volatility) was used to verify proper calibration of the instrument. The reagent was subjected to the same test conditions as the samples [7]. Moreover, the instrument calibration was checked daily using the reagent 2,3-dimethylbutane, according to ASTM D 5191, that has a similar vapor pressure value to Brazilian gasoline (57.1 ± 0.2 °C). 2.3. Experimental procedures 2.3.1. Distillation curves The samples were analyzed in a Herzog HDA 627 automatic distiller according to the ASTM D86 standard [18]. For the distillation test, 100 mL of previously cooled gasoline was transferred to a specific distillation flask coupled to a sensor and heated to maintain the distillation rate between 4 and 5 mL min1. The distilled vapor was condensed and collected in a cooled beaker and the distillation curves (distillation temperature versus volume recovered), at 1% (v/v) intervals, were obtained after correcting temperature readings at atmospheric pressure to 760 mmHg and considering volume loss, according to ASTM D86 [18]. 2.3.2. Manual distillation A manual distillation system adapted to the specifications established by ASTM D86 was used for the analysis of distillation fractions of automotive fuel [18]. The samples were prepared by adding 25% (v/v) ethanol to regular gasoline, originally from the REGAP refinery. Most of the samples were originally from this refinery. The specifications used for the manual method were the same as those for the automatic method, with distilled fractions being collected every 10 mL. Five distillations of this gasoline were carried out, and each distillation fraction was mixed to increase the representation of the results. The solutions of the different percentages of recovered volume were cooled (from 8 to 15 °C) and later submitted to spectrometric analysis in a commercial infrared spectrometer to determine the composition of the distilled fractions.
G. Mendes et al. / Fuel 187 (2017) 167–172
2.3.3. Gas Chromatography–Mass Spectrography (GC–MS) The composition of the distilled fractions was obtained using electron-ionization mass spectrometry (EI-MS). The analyses were performed in a Shimadzu model GC-17A/QP-5050A GC–MS, using a fused capillary column (50 m 0.2 mm 0.5 lm, PONA50, HP), with poly(methyl siloxane) as the stationary phase and helium as the carrier gas at a constant flow rate of 0.1 mL min1. Sample aliquots of 1.0 lL were injected in split mode (1:16) without solvent delay. The analyses were carried out in the following conditions: initial temperature 34 °C for eight minutes, then increase at 2 °C per minute until reaching 60 °C, 3 °C per minute up to 185 °C and 10 °C per minute up to 250 °C, where the temperature was maintained for 2 min. Injector and detector temperatures were 230 and 250 °C, respectively. The mass spectrometer worked in ionization mode of 70 eV and scan mode (m/z 45–350). The presence of different types of compounds in the samples was discovered using total ion chromatograms (TIC) together with library information (Wiley Class 5000, 6th edition). Compounds with less than 90% similarity of mass spectra were discarded. 2.3.4. Infrared analysis The concentrations of paraffins, olefins and aromatics were obtained using an automatic PetroSpec GS1000 analyzer based in medium infrared spectroscopy combined with the multivariate calibration method. The data were analyzed by MLR (multiple linear regression), according to ASTM E1655 [19]. The device’s database contains chromatographic data on standard samples and is updated periodically with the introduction of new samples [20]. The accuracy of the values obtained with this spectrometer is guaranteed by routine evaluation of samples for Interlaboratory Program of the ANP, which has more than 20 participants and has at least two annual testing rounds [3]. 2.4. Statistical analysis In the multivariate regression model, the distillation curves of gasoline samples were arranged in an array X, while the reference vapor pressure values (obtained from the assay based on ASTM D5191) were allocated in a vector y. For correlation between the distillation curves (X) and the reference vapor pressure values (y), we used the Minitab software (release 14 for Windows) and Matlab 2011b (MathWorks, Natick, MA, USA) using PLS Toolbox 7.5 (Eigenvector Research, Wenatchee, WA, USA). With regression by partial least squares, the arrays X and y are decomposed into smaller matrices that contain relevant information about the samples (called scores) and the original variables (loadings).
X ¼ TPt þ E
ð1Þ
y ¼ Tqt þ f
ð2Þ
where T is the scores matrix, P is the loadings matrix for X and q is the loading vector for y. E and f are the residuals representing the non-modeled portion of the data. After this matrix decomposition, a linear relationship is established between the scores of X and y, generating a regression model. For the prediction of new samples ÿ can be used Eq. (3).
€ ¼ XbPLS y
ð3Þ
where bPLS is the regression coefficient. From this linear relationship it is possible to determine values of the property of interest, analyze similarities, groupings and outliers from the chart of scores, and identify original X variables with greater importance in the pre-
169
diction of y, by evaluating the loadings chart. The higher the value of the factor loading, the more important the variable is [21]. For this study, the gasoline samples were divided into two sets (calibration and external validation) with 53 and 27 samples, respectively, using the Kennard-Stone algorithm [22]. For data preprocessing, autoscaling was used to assign equal importance to the variables, as was made in others studies using distillation curves [11–17]. For internal validation, the ‘‘leave-one-out” method was used while the number of latent variables in the calibration set was selected based on the lowest PRESS value (prediction error sum of squares) [23]. The appropriate number of variables was selected using the F-test for each pair of PRESS values. The accuracy was obtained by calculating the average error estimate of the actual vapor pressure, compared to the prediction results obtained by root mean square error prediction (RMSEP). In addition, we calculated the correlation coefficient (R) between the reference and predicted values to evaluate the data adjustment of the calibration and external validation models. The test was also applied to tbias samples from external validation to investigate if the systematic errors found in the model were significant. The evaluation of repeatability and reproducibility of the methods was carried out according to ISO-5725-2 [24]. Therefore, for the Reid vapor pressure, seven samples of gasoline were used and, for each sample, assays were done by three different analysts with seven replicates for each one [25], producing a total of 147 results. Anomalous samples were detected by comparing the values of residue and the leverage of data sets. A sample is considered anomalous when it simultaneously presents high leverage and residual values [21]. No anomalous samples were noted in the dataset used in the prediction of Reid vapor pressure.
3. Results and discussion For prediction of vapor pressure, a calibration model was constructed where the data matrix X and reference values y were treated with the PLS regression method. The number of latent variables used in the PLS model was determined from the PRESS values displayed in the calibration set, with the smallest value, after an F-test, being used to construct the models [23]. Fig. 1 shows the PRESS values obtained in the calibration set for the autoscaled data, which led to the choice of eight latent variables. This choice was intended to prevent overfitting of the model, since the PRESS values did not show statistically significant difference at 95% from the F-test using more variables. The eight latent variables explained 97.9% of variance in X and 96.8% in y. As the other distillation curves studies [11–17,26], the preprocessing data was tested with the original data (autoscaled and mean centered) and the RMSEC values obtained in each case are shown in Table 1. Autoscaling was the selected preprocessing for the subsequent studies, since the RMSEC value obtained was the smallest among the values obtained using data centered on the mean and also data without preprocessing. The F-test showed there was a significant difference between the RMSEC values obtained from mean-centered and autoscaled data as well as the values of the original autoscaled data obtained by RMSEC. The same test indicated there was no significant difference between the RMSEC values obtained when the data were mean centered and not preprocessed. The same result was obtained in other works using distillation curves and PLS, a characteristic of distillation curves [11–17,26]. Thus, as in other studies using distillation curves to predict physico-chemical properties of gasoline, the preprocessing used in prediction of vapor pressure was autoscaling because it presented the smallest RMSEC value. After determining the best conditions for construction of the model and verifying the absence of anomalous samples, we carried
170
G. Mendes et al. / Fuel 187 (2017) 167–172
Fig. 1. PRESS values as function of number of latent variables in determination of vapor pressure using distillation curves (ASTM D86).
Table 1 RMSEC values for different types of preprocessing to predict automotive gasoline vapor pressure. Preprocessing
RVP (kPa)
Mean centering data Autoscaled Original data
0.92 0.71 1.02
out a study of the important variables in predicting the vapor pressure through the loadings chart (Fig. 2). This figure indicates that in the first latent variable the most important fractions cover the range of 4–60% (v/v), precisely the most volatile part of the fuel. Fig. 3 shows the values of vapor pressure of each fraction of a distilled gasoline containing 25% (v/v) ethanol. This sample was distilled and each of the fractions was submitted to vapor pressure analysis. The highest vapor pressure values occurred in the initial fractions, after which they decreased progressively during the distillation. This indicates that the initial fractions present greater levels of light hydrocarbons, which have higher vapor pressure values. This was confirmed by analyzing the same fractions through infrared spectrometry and chromatography. The composition of the distilled fractions obtained by infrared showed that levels of paraffins were higher than the olefins and ethanol, indicating that paraffins have the strongest influence on the fractions’ vapor pressure values (Fig. 4), mainly the 60% (v/v) fraction. The final fractions consisted of the aromatic hydrocarbons, which have greater influence. Hydrocarbons such as hexane
Fig. 3. Vapor pressure values for the various distillation fractions (ASTM D86) of a gasoline containing 25% (v/v) ethanol.
have high vapor pressure values and low boiling temperatures [27]. In contrast, aromatic hydrocarbons such as toluene have low vapor pressure values and high boiling points [27]. The results obtained by gas chromatography confirmed the previous findings, indicating that the final fractions are rich in aromatic hydrocarbons (ethyl benzene, 1,2,4-trimethylbenzene, 1.4dietilbenzene) and the initial fractions are rich in other hydrocar-
Fig. 2. Loading graph for first latent variables for the determination of vapor pressure using distillation curves (ASTM D86).
G. Mendes et al. / Fuel 187 (2017) 167–172
171
values. Fig. 5 shows the fit between these values for samples from the calibration set and for samples from the validation set. The fit obtained (0.9840) was better than the one found by Flecher et al. in determining the vapor pressure of automotive gasoline using Raman spectroscopy [9]. The different origins of the samples used in this study did not prevent obtaining low RMSEP values compared to models in the literature, and it is possible to obtain high correlation between the reference and predicted values.
4. Conclusion
Fig. 4. Percentage of different hydrocarbons in distilled fractions (ASTM D86) of automotive gasoline with 25% (v/v) ethanol. (h) Olefins, (.) aromatics, (d) ethanol (D) paraffins.
bons such as hexane, methylcyclopentane, 3-methyl-1-pentene, 2methyl-1-pentene and cyclopentane. After choosing the appropriate number of latent variables and the best preprocessing method, another important step is to evaluate the accuracy of the proposed method, to have a model able to predict new samples. The evaluation of the quality of the models was done by calculating some parameters as RMSEP, bias (using t test as 95% of confidence), repeatability and reproducibility (Table 2). The RMSEP value obtained by the proposed method for the prediction of vapor pressure of automotive gasoline, 0.67 kPa, was less than the values described in the literature using other analytical techniques, as Medium Infrared (2.81) [9], Raman (3.92) [7], and Near Infrared (5.99) [8]. This result indicates good performance of the model used in spite of the different profiles in the distillation curves caused by the varied origins of the samples, demonstrating the accuracy of the proposed method. The t-test was used to detect the presence of systematic errors (bias) [19]. The t-value calculated (tcalc) was lower (2.74 1014 kPa) than the tabled value (ttab 26 FD = 2.07), with a confidence level of 95%, showing that the systematic error in the model can be considered insignificant and disregarded. Table 2 presents the repeatability and reproducibility values for the Reid vapor pressure assay of the proposed method, also the standard method [7], as well as the maximum permitted values for this assay. In both parameters obtained from the proposed method, the values were less than the maximum values permitted by ASTM D5191, indicating that the proposed methodology is highly accurate. To evaluate the fit of the data, the vapor pressure values provided by the proposed method were compared to the reference
The use of distillation curves in conjunction with multivariate calibration PLS was able to make predictions of one of the most important properties related to automotive gasoline volatility, vapor pressure, with values in the interval of 49.1–67.4 kPa, using samples from different sources, providing greater reliability to the model. We obtained a good degree of agreement between the value estimated by the multivariate model and reference (true) value. The results were better than those obtained by other analytical techniques reported in the literature. The use of distillation curves provided a model for the prediction of the vapor pressure of gasoline with excellent performance and low RMSEP value compared with the models obtained from spectroscopic techniques [8–11], despite the different profiles in the distillation curves caused by the varied origin of the samples. We obtained good adjustment of reference and predicted values both for the results of calibration (R = 0.9840) and validation (R = 0.9743) for determination of the vapor pressure using PLS and distillation curves. These results were better than those obtained by Flecher et al. [9]. The good fit obtained indicates that the models built from distillation curves are efficient with low dispersion of points along the line by the least squares method. The analysis of loadings graphs indicated that the initial distillation curve variables used in this study were important for the construction of the PLS model according to the composition determined by gas chromatography and infrared spectroscopy. The proposed method for determination of vapor pressure from distillation curves and multivariate calibration proved useful in speeding up the analytical process and reducing the cost of analysis. This is because the distillation test (present in the scope of ANP analysis for assessing the quality of automotive gasoline) is already
Table 2 RMSEP values, among other parameters, obtained in the prediction of RVP using PLS multivariate calibration. Parameter
RVP (kPa)
RMSEP t-test (tcalc) validation set t-test (ttab) validation set Repeatability (proposed method) Reproducibility (proposed method) Repeatability (ASTM D5191) Reproducibility (ASTM D5191)
0.67 2.74 1014 2.07 0.65 0.91 1.47 2.45
Fig. 5. Reference value versus predicted values for prediction of vapor pressure, for the calibration set (d) and validation set (s).
172
G. Mendes et al. / Fuel 187 (2017) 167–172
routinely performed, which makes it not necessary to carry out a new analysis (for example, an NIR spectrum) to predict RVP. Acknowledgements The authors are thankful to FINEP – Brazil, FAPEMIG – Brazil and CNPq – Brazil for the financial support. References [1] Campos AC, Leontsinis E. Petróleo & Derivados: obtenção, especificações, requisitos e desempenho. São Paulo: Editora Técnica Ltda; 1990. [2] Neiva J. Conheça o Petróleo. Rio de Janeiro: Expressão e Cultura Editora; 1993. [3] Annual Data of Brazilian Fuel Sales, at [accessed May 2016]. [4] Takeshita EV, Rezende RVP, de Souza SMAGU, de Souza AAU. Influence of solvent addition on the physicochemical properties of Brazilian gasoline. Fuel 2008;87:2168–77. [5] ANP – Agência Nacional do Petróleo, Gás Natural e Biocombustíveis; Resolução no 40 de 25.10.2013. [6] Cataluña R, Silva R. Desenvolvimento de um equipamento para avaliação do efeito do etanol na pressão de vapor e entalpia de vaporização em gasolinas automotivas. Quim Nova 2006;29:580–5. [7] ASTM – American Society for Testing and Materials – Standard Test Method for Vapor Pressure of Petroleum Products (Mini Method). ASTM D5191, West Conshohocken PA; 2013. [8] Cooper JB, Wise KL, Groves J, Welch W. Determination of octane number and Reid vapor pressure of commercial petroleum fuels using FT-Raman spectroscopy and partial least-squares regression analysis. Anal Chem 1995;67:4096–100. [9] Flecher PE, Welch WT, Albin S, Cooper JB. Determination of octane numbers and Reid vapor pressure in commercial gasoline using dispersive fiber-optic Raman spectroscopy. Spectrochim Acta A 1997;53:99–206. [10] Côcco LC, Yamamoto CI, von Meien OF. Study of correlations for physicochemical properties of Brazilian gasoline. Chemometr Intell Lab 2005;76:55–63. [11] Barbeira PJS, Aleme HG, Costa LM. Determination of ethanol and specific gravity in gasoline by distillation curves and multivariate analysis. Talanta 2009;78:1422–8.
[12] Barbeira PJS, Aleme HG, Mendes G. Determination of octane numbers in gasoline by distillation curves and partial least squares regression. Fuel 2012;97:131–6. [13] Barbeira PJS, Aleme HG. Determination of flash point and cetane index in diesel using distillation curves and multivariate calibration. Fuel 2012;102:129–34. [14] Barbeira PJS, Aleme HG. Determination of biodiesel content in diesel using distillation curves and multivariate calibration. Energy Fuels 2012;26:5769–74. [15] Barbeira PJS, Aleme HG, Assunção RA, Carvalho MMO. Determination of specific gravity and kinematic viscosity of diesel using distillation curves and multivariate calibration. Fuel Process Technol 2012;102:90–5. [16] Barbeira PJS, Aleme HG, Costa LM. Determination of gasoline origin by distillation curves and multivariate analysis. Fuel 2008;87:3664–8. [17] Barbeira PJS, Mendes G. Detection and quantification of adulterants in gasoline using distillation curves and multivariate methods. Fuel 2013;112:163–71. [18] ASTM – American Society for Testing and Materials – Standard Test for Distillation of Petroleum Products at Atmospheric Pressure. ASTM D86, West Conshohocken PA; 2012. [19] ASTM – American Society for Testing and Materials – Standard Practices for Infrared Multivariate Quantitative Analysis. ASTM E1655, West Conshohocken PA; 2012. [20] Petrospec Calibration Update Software Windows Version Operating Manual Version 1 May-1999. [21] Massart DL, Vandeginste BGM, Buydens LMC, Jong S, Lewi PJ, Smeyers-Verbeke J. Handbook of chemometrics and qualimetrics. Part B. Amsterdam: Elsevier Publ; 1998. [22] Kennard RW, Stone LA. Computer aided design of experiments. Technometrics 1969;11:137–48. [23] Geladi P, Kowalsky BR. Partial least-squares regression: a tutorial. Anal Chim Acta 1986;185:1–17. [24] ISO/IEC 5725–2:2002 Accuracy (Trueness and precision) of measurement methods and results. Part 2. Basic method for the determination of repeatability and reproducibility of a standard measurement method. [25] INMETRO DOQ-CGCRE-008, Orientação sobre Validação de Métodos de Ensaios Químicos; 2010. [26] Barbeira PJS, Aleme HG, Corgozinho CNC. Diesel oil discrimination by origin and type using physicochemical properties and multivariate analysis. Fuel 2010;89:3151–6. [27] Barbosa LCA. Introdução à Química Orgânica. São Paulo: Pearson; 2011.
Contents lists available at ScienceDirect
Fuel journal homepage: www.elsevier.com/locate/fuel
Full Length Article
Reid vapor pressure prediction of automotive gasoline using distillation curves and multivariate calibration Gisele Mendes a, Helga G. Aleme b, Paulo J.S. Barbeira a,⇑ a Laboratório de Ensaios de Combustíveis, Departamento de Química, ICEx, Universidade Federal de Minas Gerais, Av. Antônio Carlos 6627, 31270-901 Belo Horizonte, Minas Gerais, Brazil b Departamento de Ciências Exatas e da Terra, ICAQF, Universidade Federal de São Paulo, R. São Nicolau 210, 09913-030 Diadema, São Paulo, Brazil
a r t i c l e
i n f o
Article history: Received 15 June 2016 Received in revised form 26 August 2016 Accepted 12 September 2016
Keywords: Vapor pressure Gasoline PLS Distillation
a b s t r a c t Partial least squares regression (PLS) in conjunction with distillation curves (ASTM D86) were used successfully to predict the vapor pressure of automotive gasoline. The errors obtained in the calibration and validation models (0.71 and 0.69 kPa) were lower than those reported in the literature. In addition, the proposed method is inexpensive, reduces test time and is easy to implement, making it an alternative method for gasoline quality control. It does be possible because the distillation tests are already routinely performed by ANP as one of the evaluation parameter of the automotive gasoline quality. Ó 2016 Elsevier Ltd. All rights reserved.
1. Introduction Automotive gasoline is a flammable liquid formulated by mixing hydrocarbons ranging from four to twelve carbon atoms, with boiling points between 30 and 220 °C, obtained by refining petroleum, involving direct distillation, cracking, reforming, isomerization and alkylation [1,2]. Its composition varies according to the method employed in its production and the chemical composition of the crude oil. In 2015 about 44.2 billion liters was sold in Brazil [3]. In this context, there is an increasing need for quality control of gasoline and the development of new technologies to ensure that the product meets the requirements of engines and does not produce emissions above legal levels [4]. According to the ANP (National Petroleum, Natural Gas and Biofuels Agency), various parameters are used in quality control of gasoline, such as distillation parameters, specific gravity, and octane numbers, among others [5]. Currently, the analyses of these parameters are carried out in Brazil using the ASTM (American Society for Testing and Materials) or NBR (Brazilian Association of Technical Standards) standard methods. A fuel must meet all the specifications to be considered of good quality. An important physical-chemical parameter in assessing the quality of gasoline is the vapor pressure, because it is related to the fuel’s volatility, which directly influences engine performance. ⇑ Corresponding author. E-mail address: [email protected] (P.J.S. Barbeira). http://dx.doi.org/10.1016/j.fuel.2016.09.046 0016-2361/Ó 2016 Elsevier Ltd. All rights reserved.
In the production of fuel, vapor pressure is measured to indicate the requirements that must be met for the transport and storage of the product, in order to avoid accidents and minimize losses by evaporation. Fuels with high vapor pressure emit more volatile compounds and, after production and shipping fuel to service stations, this test can be applied to evaluate nonconformities related to adulteration caused by adding solvents. Takeshita et al. [4] noted that the addition of organic solvents at different ratios modifies the vapor pressure of gasoline, an effect that is more preponderant at 50% (v/v) and above. The evaluation of vapor pressure is also required to assure proper engine performance, since fuels with low vapor pressures are harder to vaporize in the intake manifold, hindering the combustion process. On the other hand, fuels with high vapor pressure emit more volatile organic compounds, which can cause incomplete fuel vaporization and uncontrolled air/fuel ratio. Hence, the vapor pressure must be within an acceptable range [6]. In Brazil, ANP establishes a maximum of 69.0 kPa for the vapor pressure, at 37.8 °C, of automotive gasoline [5]. The volatility parameters, which affect the performance at low temperature, are typically determined by testing vehicles at high RPM. At high temperatures, the most common phenomena that affect fuel systems in vehicles are vapor lock and percolation. Vapor lock is the occurrence of a mass of vapor between the fuel tank and carburetor or fuel injection system. Percolation is the result of uncontrolled fuel vaporization, which can happen after a long period operating at high temperatures [6].
168
G. Mendes et al. / Fuel 187 (2017) 167–172
Vapor pressure can be measured following ASTM D5191 [7]. The test conditions are carefully specified since the vapor pressure varies with temperature, the amount of dissolved air and the ratio between vapor and liquid in the container. In this method, the sample contained in a jar (filled to between 70 and 80% of capacity) is cooled to temperatures between 0 and 1 °C and subsequently inserted into the liquid chamber and the equipment is connected to the vapor chamber heated to 37.8 °C. The increased pressure in the vapor chamber is measured using a sensor transducer and a pressure indicator and the vapor pressure is registered on the display. This measure is defined as Reid vapor pressure (RVP) and can be expressed in kPa or psi. Although the test based on ASTM D5191 be relatively quick (approximately seven minutes), it is desirable to have simpler and low cost methods for RVP determination. In this context, chemometric methods combined with distillation curves can provide important information about different gasoline quality parameters, reducing the number of tests to being performed and used as screening for more specific tests. The current literature demonstrates the predictive power of multivariate regression models for vapor pressure prediction of automotive gasoline [8–11]. Among the regression techniques available, the most widely used are partial least squares regression (PLS) and principal components regression (PCR). Cooper et al. [8] correlated the results of vapor pressure obtained experimentally together with the Raman spectra of fuels to build a regression model using PLS. The authors found RMSEP values (root mean squares error of prediction) of 3.92 kPa. Using dispersive fiber-optic Raman spectroscopy with a CCD detector and near infrared, Flecher et al. [9] determined the vapor pressure of gasoline samples, finding RMSEP of 5.99 kPa. Besides PLS, Côcco et al. [10] associated near-infrared spectroscopy and the chemometric tool PCR to determine different chemical and physical parameters, including vapor pressure. For this property, the value of RMSEP obtained was 2.81 kPa. Recently, several works have demonstrated the good potential of distillation curves in conjunction with PLS regression to forecast quality parameters, such as specific gravity and ethanol content [11], MON and RON in automotive gasoline [12], flash point [13], cetane number [13], biodiesel content [14], specific gravity [15] and kinematic viscosity [15] in diesel oil. In addition to these parameters, these models have the ability to discriminate the refinery of origin [16] and to detect adulteration by solvents [17]. Distillation is a method of separation based on the phenomenon of vapor-liquid equilibrium for mixtures. With the ASTM D86 method it is possible to evaluate the complexity of liquid blends, which is directly related to the volatility of the sample’s components [18]. To simplify the analysis of gasoline through the use of a single test, thus reducing costs and increasing the number of samples that can be analyzed, this work describes the use of distillation curves from regular tests to evaluate the quality of gasoline, obtained according to ASTM D86 [18], in conjunction with multivariate calibration PLS for determination of the vapor pressure. We carried out manual distillations to obtain the composition of each fraction using infrared spectrometry with the purpose of showing the most important variables in predicting vapor pressure.
2. Experimental 2.1. Samples To build the PLS model used in the determination of the vapor pressure of gasoline, 80 samples of premium and regular gasolines
(both containing ethanol 25 % v/v) were obtained from service stations in the eastern region of the state of Minas Gerais (Brazil), which receives fuels produced by five different refineries. A study proposed by Aleme et al. [16] showed that gasolines produced by these refineries have different distillation curve profiles, due to their different chemical composition, related to the source of crude oil and the refining process thereof. The samples were stored in polyethylene bottles, sealed and chilled to 8–15 °C to avoid evaporation of volatile components. All samples were subjected to the tests of physico-chemical parameters established by the ANP [5]. 2.2. Materials and equipment 2.2.1. Vapor pressure analysis Before starting the tests, the samples were transferred to 50 mL flasks, filled to between 70 and 80% capacity and sealed. The samples were cooled in an ice-water bath to a temperature between 5 and 7 °C. After reaching the proper temperature, the flasks were shaken and opened momentarily. This process was repeated three times. Before the readings, the measuring chamber was cleaned with toluene and dried with n-pentane with the aid of a vacuum pump. After ensuring that the measuring chamber was completely dry, one of the valves was closed and a volume of 2.55 mL of air was introduced with a syringe, after which the valve was closed quickly to achieve pressure between 19.0 and 21.0 kPa. After one minute for stabilization, 2.05 mL of sample was and injected into the measuring chamber, with temperature of 37.8 ± 0.1 °C, using a syringe. The results were displayed in approximately three minutes, in kPa. The test was repeated three times for each sample. The reagent npentane (known volatility) was used to verify proper calibration of the instrument. The reagent was subjected to the same test conditions as the samples [7]. Moreover, the instrument calibration was checked daily using the reagent 2,3-dimethylbutane, according to ASTM D 5191, that has a similar vapor pressure value to Brazilian gasoline (57.1 ± 0.2 °C). 2.3. Experimental procedures 2.3.1. Distillation curves The samples were analyzed in a Herzog HDA 627 automatic distiller according to the ASTM D86 standard [18]. For the distillation test, 100 mL of previously cooled gasoline was transferred to a specific distillation flask coupled to a sensor and heated to maintain the distillation rate between 4 and 5 mL min1. The distilled vapor was condensed and collected in a cooled beaker and the distillation curves (distillation temperature versus volume recovered), at 1% (v/v) intervals, were obtained after correcting temperature readings at atmospheric pressure to 760 mmHg and considering volume loss, according to ASTM D86 [18]. 2.3.2. Manual distillation A manual distillation system adapted to the specifications established by ASTM D86 was used for the analysis of distillation fractions of automotive fuel [18]. The samples were prepared by adding 25% (v/v) ethanol to regular gasoline, originally from the REGAP refinery. Most of the samples were originally from this refinery. The specifications used for the manual method were the same as those for the automatic method, with distilled fractions being collected every 10 mL. Five distillations of this gasoline were carried out, and each distillation fraction was mixed to increase the representation of the results. The solutions of the different percentages of recovered volume were cooled (from 8 to 15 °C) and later submitted to spectrometric analysis in a commercial infrared spectrometer to determine the composition of the distilled fractions.
G. Mendes et al. / Fuel 187 (2017) 167–172
2.3.3. Gas Chromatography–Mass Spectrography (GC–MS) The composition of the distilled fractions was obtained using electron-ionization mass spectrometry (EI-MS). The analyses were performed in a Shimadzu model GC-17A/QP-5050A GC–MS, using a fused capillary column (50 m 0.2 mm 0.5 lm, PONA50, HP), with poly(methyl siloxane) as the stationary phase and helium as the carrier gas at a constant flow rate of 0.1 mL min1. Sample aliquots of 1.0 lL were injected in split mode (1:16) without solvent delay. The analyses were carried out in the following conditions: initial temperature 34 °C for eight minutes, then increase at 2 °C per minute until reaching 60 °C, 3 °C per minute up to 185 °C and 10 °C per minute up to 250 °C, where the temperature was maintained for 2 min. Injector and detector temperatures were 230 and 250 °C, respectively. The mass spectrometer worked in ionization mode of 70 eV and scan mode (m/z 45–350). The presence of different types of compounds in the samples was discovered using total ion chromatograms (TIC) together with library information (Wiley Class 5000, 6th edition). Compounds with less than 90% similarity of mass spectra were discarded. 2.3.4. Infrared analysis The concentrations of paraffins, olefins and aromatics were obtained using an automatic PetroSpec GS1000 analyzer based in medium infrared spectroscopy combined with the multivariate calibration method. The data were analyzed by MLR (multiple linear regression), according to ASTM E1655 [19]. The device’s database contains chromatographic data on standard samples and is updated periodically with the introduction of new samples [20]. The accuracy of the values obtained with this spectrometer is guaranteed by routine evaluation of samples for Interlaboratory Program of the ANP, which has more than 20 participants and has at least two annual testing rounds [3]. 2.4. Statistical analysis In the multivariate regression model, the distillation curves of gasoline samples were arranged in an array X, while the reference vapor pressure values (obtained from the assay based on ASTM D5191) were allocated in a vector y. For correlation between the distillation curves (X) and the reference vapor pressure values (y), we used the Minitab software (release 14 for Windows) and Matlab 2011b (MathWorks, Natick, MA, USA) using PLS Toolbox 7.5 (Eigenvector Research, Wenatchee, WA, USA). With regression by partial least squares, the arrays X and y are decomposed into smaller matrices that contain relevant information about the samples (called scores) and the original variables (loadings).
X ¼ TPt þ E
ð1Þ
y ¼ Tqt þ f
ð2Þ
where T is the scores matrix, P is the loadings matrix for X and q is the loading vector for y. E and f are the residuals representing the non-modeled portion of the data. After this matrix decomposition, a linear relationship is established between the scores of X and y, generating a regression model. For the prediction of new samples ÿ can be used Eq. (3).
€ ¼ XbPLS y
ð3Þ
where bPLS is the regression coefficient. From this linear relationship it is possible to determine values of the property of interest, analyze similarities, groupings and outliers from the chart of scores, and identify original X variables with greater importance in the pre-
169
diction of y, by evaluating the loadings chart. The higher the value of the factor loading, the more important the variable is [21]. For this study, the gasoline samples were divided into two sets (calibration and external validation) with 53 and 27 samples, respectively, using the Kennard-Stone algorithm [22]. For data preprocessing, autoscaling was used to assign equal importance to the variables, as was made in others studies using distillation curves [11–17]. For internal validation, the ‘‘leave-one-out” method was used while the number of latent variables in the calibration set was selected based on the lowest PRESS value (prediction error sum of squares) [23]. The appropriate number of variables was selected using the F-test for each pair of PRESS values. The accuracy was obtained by calculating the average error estimate of the actual vapor pressure, compared to the prediction results obtained by root mean square error prediction (RMSEP). In addition, we calculated the correlation coefficient (R) between the reference and predicted values to evaluate the data adjustment of the calibration and external validation models. The test was also applied to tbias samples from external validation to investigate if the systematic errors found in the model were significant. The evaluation of repeatability and reproducibility of the methods was carried out according to ISO-5725-2 [24]. Therefore, for the Reid vapor pressure, seven samples of gasoline were used and, for each sample, assays were done by three different analysts with seven replicates for each one [25], producing a total of 147 results. Anomalous samples were detected by comparing the values of residue and the leverage of data sets. A sample is considered anomalous when it simultaneously presents high leverage and residual values [21]. No anomalous samples were noted in the dataset used in the prediction of Reid vapor pressure.
3. Results and discussion For prediction of vapor pressure, a calibration model was constructed where the data matrix X and reference values y were treated with the PLS regression method. The number of latent variables used in the PLS model was determined from the PRESS values displayed in the calibration set, with the smallest value, after an F-test, being used to construct the models [23]. Fig. 1 shows the PRESS values obtained in the calibration set for the autoscaled data, which led to the choice of eight latent variables. This choice was intended to prevent overfitting of the model, since the PRESS values did not show statistically significant difference at 95% from the F-test using more variables. The eight latent variables explained 97.9% of variance in X and 96.8% in y. As the other distillation curves studies [11–17,26], the preprocessing data was tested with the original data (autoscaled and mean centered) and the RMSEC values obtained in each case are shown in Table 1. Autoscaling was the selected preprocessing for the subsequent studies, since the RMSEC value obtained was the smallest among the values obtained using data centered on the mean and also data without preprocessing. The F-test showed there was a significant difference between the RMSEC values obtained from mean-centered and autoscaled data as well as the values of the original autoscaled data obtained by RMSEC. The same test indicated there was no significant difference between the RMSEC values obtained when the data were mean centered and not preprocessed. The same result was obtained in other works using distillation curves and PLS, a characteristic of distillation curves [11–17,26]. Thus, as in other studies using distillation curves to predict physico-chemical properties of gasoline, the preprocessing used in prediction of vapor pressure was autoscaling because it presented the smallest RMSEC value. After determining the best conditions for construction of the model and verifying the absence of anomalous samples, we carried
170
G. Mendes et al. / Fuel 187 (2017) 167–172
Fig. 1. PRESS values as function of number of latent variables in determination of vapor pressure using distillation curves (ASTM D86).
Table 1 RMSEC values for different types of preprocessing to predict automotive gasoline vapor pressure. Preprocessing
RVP (kPa)
Mean centering data Autoscaled Original data
0.92 0.71 1.02
out a study of the important variables in predicting the vapor pressure through the loadings chart (Fig. 2). This figure indicates that in the first latent variable the most important fractions cover the range of 4–60% (v/v), precisely the most volatile part of the fuel. Fig. 3 shows the values of vapor pressure of each fraction of a distilled gasoline containing 25% (v/v) ethanol. This sample was distilled and each of the fractions was submitted to vapor pressure analysis. The highest vapor pressure values occurred in the initial fractions, after which they decreased progressively during the distillation. This indicates that the initial fractions present greater levels of light hydrocarbons, which have higher vapor pressure values. This was confirmed by analyzing the same fractions through infrared spectrometry and chromatography. The composition of the distilled fractions obtained by infrared showed that levels of paraffins were higher than the olefins and ethanol, indicating that paraffins have the strongest influence on the fractions’ vapor pressure values (Fig. 4), mainly the 60% (v/v) fraction. The final fractions consisted of the aromatic hydrocarbons, which have greater influence. Hydrocarbons such as hexane
Fig. 3. Vapor pressure values for the various distillation fractions (ASTM D86) of a gasoline containing 25% (v/v) ethanol.
have high vapor pressure values and low boiling temperatures [27]. In contrast, aromatic hydrocarbons such as toluene have low vapor pressure values and high boiling points [27]. The results obtained by gas chromatography confirmed the previous findings, indicating that the final fractions are rich in aromatic hydrocarbons (ethyl benzene, 1,2,4-trimethylbenzene, 1.4dietilbenzene) and the initial fractions are rich in other hydrocar-
Fig. 2. Loading graph for first latent variables for the determination of vapor pressure using distillation curves (ASTM D86).
G. Mendes et al. / Fuel 187 (2017) 167–172
171
values. Fig. 5 shows the fit between these values for samples from the calibration set and for samples from the validation set. The fit obtained (0.9840) was better than the one found by Flecher et al. in determining the vapor pressure of automotive gasoline using Raman spectroscopy [9]. The different origins of the samples used in this study did not prevent obtaining low RMSEP values compared to models in the literature, and it is possible to obtain high correlation between the reference and predicted values.
4. Conclusion
Fig. 4. Percentage of different hydrocarbons in distilled fractions (ASTM D86) of automotive gasoline with 25% (v/v) ethanol. (h) Olefins, (.) aromatics, (d) ethanol (D) paraffins.
bons such as hexane, methylcyclopentane, 3-methyl-1-pentene, 2methyl-1-pentene and cyclopentane. After choosing the appropriate number of latent variables and the best preprocessing method, another important step is to evaluate the accuracy of the proposed method, to have a model able to predict new samples. The evaluation of the quality of the models was done by calculating some parameters as RMSEP, bias (using t test as 95% of confidence), repeatability and reproducibility (Table 2). The RMSEP value obtained by the proposed method for the prediction of vapor pressure of automotive gasoline, 0.67 kPa, was less than the values described in the literature using other analytical techniques, as Medium Infrared (2.81) [9], Raman (3.92) [7], and Near Infrared (5.99) [8]. This result indicates good performance of the model used in spite of the different profiles in the distillation curves caused by the varied origins of the samples, demonstrating the accuracy of the proposed method. The t-test was used to detect the presence of systematic errors (bias) [19]. The t-value calculated (tcalc) was lower (2.74 1014 kPa) than the tabled value (ttab 26 FD = 2.07), with a confidence level of 95%, showing that the systematic error in the model can be considered insignificant and disregarded. Table 2 presents the repeatability and reproducibility values for the Reid vapor pressure assay of the proposed method, also the standard method [7], as well as the maximum permitted values for this assay. In both parameters obtained from the proposed method, the values were less than the maximum values permitted by ASTM D5191, indicating that the proposed methodology is highly accurate. To evaluate the fit of the data, the vapor pressure values provided by the proposed method were compared to the reference
The use of distillation curves in conjunction with multivariate calibration PLS was able to make predictions of one of the most important properties related to automotive gasoline volatility, vapor pressure, with values in the interval of 49.1–67.4 kPa, using samples from different sources, providing greater reliability to the model. We obtained a good degree of agreement between the value estimated by the multivariate model and reference (true) value. The results were better than those obtained by other analytical techniques reported in the literature. The use of distillation curves provided a model for the prediction of the vapor pressure of gasoline with excellent performance and low RMSEP value compared with the models obtained from spectroscopic techniques [8–11], despite the different profiles in the distillation curves caused by the varied origin of the samples. We obtained good adjustment of reference and predicted values both for the results of calibration (R = 0.9840) and validation (R = 0.9743) for determination of the vapor pressure using PLS and distillation curves. These results were better than those obtained by Flecher et al. [9]. The good fit obtained indicates that the models built from distillation curves are efficient with low dispersion of points along the line by the least squares method. The analysis of loadings graphs indicated that the initial distillation curve variables used in this study were important for the construction of the PLS model according to the composition determined by gas chromatography and infrared spectroscopy. The proposed method for determination of vapor pressure from distillation curves and multivariate calibration proved useful in speeding up the analytical process and reducing the cost of analysis. This is because the distillation test (present in the scope of ANP analysis for assessing the quality of automotive gasoline) is already
Table 2 RMSEP values, among other parameters, obtained in the prediction of RVP using PLS multivariate calibration. Parameter
RVP (kPa)
RMSEP t-test (tcalc) validation set t-test (ttab) validation set Repeatability (proposed method) Reproducibility (proposed method) Repeatability (ASTM D5191) Reproducibility (ASTM D5191)
0.67 2.74 1014 2.07 0.65 0.91 1.47 2.45
Fig. 5. Reference value versus predicted values for prediction of vapor pressure, for the calibration set (d) and validation set (s).
172
G. Mendes et al. / Fuel 187 (2017) 167–172
routinely performed, which makes it not necessary to carry out a new analysis (for example, an NIR spectrum) to predict RVP. Acknowledgements The authors are thankful to FINEP – Brazil, FAPEMIG – Brazil and CNPq – Brazil for the financial support. References [1] Campos AC, Leontsinis E. Petróleo & Derivados: obtenção, especificações, requisitos e desempenho. São Paulo: Editora Técnica Ltda; 1990. [2] Neiva J. Conheça o Petróleo. Rio de Janeiro: Expressão e Cultura Editora; 1993. [3] Annual Data of Brazilian Fuel Sales, at
[12] Barbeira PJS, Aleme HG, Mendes G. Determination of octane numbers in gasoline by distillation curves and partial least squares regression. Fuel 2012;97:131–6. [13] Barbeira PJS, Aleme HG. Determination of flash point and cetane index in diesel using distillation curves and multivariate calibration. Fuel 2012;102:129–34. [14] Barbeira PJS, Aleme HG. Determination of biodiesel content in diesel using distillation curves and multivariate calibration. Energy Fuels 2012;26:5769–74. [15] Barbeira PJS, Aleme HG, Assunção RA, Carvalho MMO. Determination of specific gravity and kinematic viscosity of diesel using distillation curves and multivariate calibration. Fuel Process Technol 2012;102:90–5. [16] Barbeira PJS, Aleme HG, Costa LM. Determination of gasoline origin by distillation curves and multivariate analysis. Fuel 2008;87:3664–8. [17] Barbeira PJS, Mendes G. Detection and quantification of adulterants in gasoline using distillation curves and multivariate methods. Fuel 2013;112:163–71. [18] ASTM – American Society for Testing and Materials – Standard Test for Distillation of Petroleum Products at Atmospheric Pressure. ASTM D86, West Conshohocken PA; 2012. [19] ASTM – American Society for Testing and Materials – Standard Practices for Infrared Multivariate Quantitative Analysis. ASTM E1655, West Conshohocken PA; 2012. [20] Petrospec Calibration Update Software Windows Version Operating Manual Version 1 May-1999. [21] Massart DL, Vandeginste BGM, Buydens LMC, Jong S, Lewi PJ, Smeyers-Verbeke J. Handbook of chemometrics and qualimetrics. Part B. Amsterdam: Elsevier Publ; 1998. [22] Kennard RW, Stone LA. Computer aided design of experiments. Technometrics 1969;11:137–48. [23] Geladi P, Kowalsky BR. Partial least-squares regression: a tutorial. Anal Chim Acta 1986;185:1–17. [24] ISO/IEC 5725–2:2002 Accuracy (Trueness and precision) of measurement methods and results. Part 2. Basic method for the determination of repeatability and reproducibility of a standard measurement method. [25] INMETRO DOQ-CGCRE-008, Orientação sobre Validação de Métodos de Ensaios Químicos; 2010. [26] Barbeira PJS, Aleme HG, Corgozinho CNC. Diesel oil discrimination by origin and type using physicochemical properties and multivariate analysis. Fuel 2010;89:3151–6. [27] Barbosa LCA. Introdução à Química Orgânica. São Paulo: Pearson; 2011.