Fuel 86 (2007) 25512559 www.fuelrst.

com

Articial neural network approaches on compositionproperty relationships of jet fuels based on GCMS
Guozhu Liu
a

a,b

, Li Wang

, Haijie Qu a,1, Huiming Shen Shuting Zhang b, Zhentao Mi a

a,*

a,1

, Xiangwen Zhang a,

Key Laboratory of Green Chemical Technology of Ministry of Education, School of Chemical Engineering and Technology, Tianjin University, Tianjin 300072, PR China b School of Environment Engineering and Technology, Tianjin University, Tianjin 300072, PR China Received 26 October 2006; received in revised form 6 February 2007; accepted 11 February 2007 Available online 19 March 2007

Abstract The relationships of compositionproperties of 80 jet fuels concerning chemical compositions and several specication properties including density, ashpoint, freezing point, aniline point and net heat of combustion were studied. The chemical compositions of the jet fuels were determined by GCMS, and grouped into eight classes of hydrocarbon compounds, including n-parans, isoparans, monocycloprans, dicycloprans, alkylbenzens, naphthalenes, tetralins, hydroaromatics. Several quantitative compositionproperty relationships were developed with three articial neural network (ANN) approaches, including single-layer feedforward neural network (SLFNN), multiple layer feedforward neural network (MLFNN) and general regressed neural network (GRNN). It was found that SLFNNs are adequate to predict density, freezing point and net heat of combustion, while MLFNNs produce better results as far as the ash point and aniline point prediction are concerned. Comparisons with the multiple linear regression (MLR) correlations reported and the standard ASTM methods showed that ANN approaches of compositionproperty relationships are signicant improvement on MLR correlations, and are comparable to the standard ASTM methods. 2007 Elsevier Ltd. All rights reserved.
Keywords: Jet fuel; Compositionproperty relations; GCMS; Articial neural network

1. Introduction The advances in aircraft systems require necessary improvements on the conventional jet fuels or development of a new generation of jet fuels to meet the rigorous requirements, such as higher thermal stability, higher heat sink, etc. [1,2]. Jet fuel, a typical middle distillate of petroleum, is a complex mixture of several 100 hydrocarbons. Generally, properties of jet fuels are limited by their physical and chemical properties to provide acceptable performance of jet engine, because compositions of jet fuels

Corresponding author. Tel.: +86 22 2789 2357; fax: +86 22 2740 2604. E-mail address: gliu@tju.edu.cn (L. Wang). The equal contribution.

vary from sample to sample as a result of variation in crude sources and rening processes. However, it is undoubted that the specied properties of jet fuels are dependent on the chemical compositions in nature. Therefore, from a practical point of view it is important to study the quantitive compositionproperty relationships and rapid quality control methods with higher reliability in order to develop the next generation of jet fuels. In the past two decades, there have been some important compositionproperty relations of jet/kerosene/diesel fuels being reported in the literatures. Cookson et al. [3] rst reported the propertycomposition relationships of 18 diesel fuels and 21 kerosene fuels, and dened the compositions by the weight percentages of n-alkanes, isoalkanes plus saturated cycloalkanes, single-ring aromatics, double-ring aromatics and polyring aromatics using HPLC

0016-2361/$ - see front matter 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.fuel.2007.02.023

2552

G. Liu et al. / Fuel 86 (2007) 25512559

Nomenclature ai BR Car Cbc Cn Ciso Ccycl D Hnc ki P Ta T coecient of multiple linear equation boiling range of fuel (C) mass fraction of aromatics mass fraction of cycloalkanes plus iso-parans mass fraction of n-parans mass fraction of iso-parans mass fraction of cyclics density (g/cm3) net heat of combustion (MJ/kg) coecient of multiple linear equation jet fuel property aniline point (C) ash point (C) Tfr T10 T90 freezing point (C) temperature at 10% distillates (C) temperature at 90% distillates (C)

Abbreviation ANN articial neural network SLFNN single-layer feedforward neural network GRNN general regressed neural network MAE mean of absolute error MARE mean of absolute relative error MLFNN multiple-layer feedforward neural network MLR multiple linear regression SD standard deviation

and 13C NMR. In the following publications, Cookson et al. [47] and Cookson and Smith [8,9] further studied the chemical compositions of more than 100 jet fuels and diesel fuels with GC, 1H NMR and 13C NMR, and found that the correlations of the properties of the fuels with fractions of aromatic carbon (Car) and n-alkyl carbon (Cn) by quantitative 13C NMR were as eective as the complex models based on HPLC + NMR or HPLC + GC, and that it was crucial to incorporate the boiling ranges into the correlations to obtain a more general compositionproperty relationships. They also provided several simple linear propertycomposition relationships taking boiling point ranges and dierent hydrocarbon classes into considerations. Caswell et al. [10] separated the fuel compositions in term of 13 carbon type groups using LC/1H NMR, which were correlated to the 17 fuel properties with an approximately equal reproducibility to the ASTM method. Fodor and Kohl [11] developed a new analytical method based on the spectra in the 4000600 cm1 wavenumber region of FTIR to measure the compositions of middle distillate fuels, and correlated the fuel properties, including heat of combustion, density, viscosity and cetane index, etc., with contents of dierent hydrocarbon classes represented by spectral intensities at dierent spectral regions. Gomez-Carracedo et al. [12] applied FTIR spectra of generated vapor-phase to characterize the kerosene (aviation jet fuel) compositions, and employed partial least square regression method to predict eight physicochemical properties (distillation curve, ash point, freezing point, percentage of aromatics and viscosity). Obviously, both NMR and FTIR only provide the qualitative knowledge about the carbon types and the chemical bonds of the fuels, but not the detail information of compounds in the fuels, which can be obtained from GCMS. In addition, a new mathematic approach should be adopted to obtain more general and accurate compositionproperty correlations, rather than the simple linear regressions. Articial neural networks (ANN) have recently attracted considerable attention in petroleum and chemical

engineering for handling complex and nonlinear problems [1316]. Yang et al. [17] presented the compositions of 69 diesel fuels in term of twelve hydrocarbon groups determined by LC and GCMS, and applied two general regression neural networks (GRNN) to correlate cetane number and density to 12 hydrocarbon groups with mean absolute errors less than 3%. Recently, Yang et al. [18] used GC-FIMS and PIONA to determine 12 chemical composition groups, including isoparans, n-parans, monocycloparans, dicycloparans, polycycloparans, alkylbenzenes, benzocycloalkanes, benzodicycloalkanes, diaromatics, triaromatics, tetraaromatics and aromatic sulfurs, for diesel fuel (200343 C), and established neural network correlations for density and cetane number with signicant improvement on the ANN correlations reported by Yang et al. [17]. However, one of the crucial properties of fuels, boiling point range, was not considered in their models. Up to now, little information has been reported on the predictions of other properties of jet fuels with this approach. The objective of this work is to present several ANN compositionproperty correlations for jet fuels, incorporating the inuences of isoparan contents and boiling ranges. GCMS was used to determine the chemical compositions, divided into 8 hydrocarbon compound classes, such as n-parans, isoparans, monocycloprans, dicycloprans, alkylbenzens, naphthalenes, tetralins, hydroaromatics. Attempts were also made to compare the performance of ANN approaches with multiple linear regression (MLR) correlations reported in the literatures and developed in this work.

2. Experimental section At rst, a compositionproperty database of 80 jet fuels was established containing chemical compositions and several specication properties (density, ashpoint, freezing point, aniline point, and net heat of combustion). Each of the jet fuels was carefully selected to ensure dierent

G. Liu et al. / Fuel 86 (2007) 25512559

2553

compositions and properties, but with the boiling point range from 180 to 280 C. Recently, GCMS was successfully applied to quantitatively analyze aromatics, n-parans, isoparans, and cycloparans of the diesel/jet fuels in the literatures [1721]. In this work, the chemical composition analyses of fuels were conducted on Agilent 6890N/5975 inert GC MSD equipped with a HP-5 MS column (30 m 0.25 mm 0.25 lm) and an electron impact (EI) ionization. High pure helium (99.999%) was used as a carrier gas with a ow rate of 1 ml/min in the column. Inlet temperature was kept at 280 C. The oven temperature was rstly held at 50 C for 2 min and then heated at 2 C/min to 260 C with a nal isothermal period of 10 min. The transfer line temperature was 280 C and Quadrupole MSD temperature was 150 C. The mass range and the scan rate were 50 350 amu and 5.46 scans/s, respectively. Fig. 1 presents a typical total ion chromatography (TIC) of a jet fuel sample used in this work. Peak purity check indicates that the normal- and isoparans are separated satisfactorily. After mass percentage of the major compounds were determined from GCMS all the compounds are grouped into eight hydrocarbon classes (see Table 1). Hydroaromatics including tetralins, alkyltetralins, decalin, alkyldecalin, etc., are classied into a class due to their lower freezing point than other bicyclics and their particularity in the supercial thermal cracking [22]. The maximum errors on the mass of hydrocarbon groups in the saturate fraction and aromatic fraction are less than 2%, as compared with standard method of AM-S 140.31. All the hydrocarbon classes described above were used as neural network inputs and independent variables in the linear regression equations. Several important jet fuel properties including density, freezing point, ashpoint, aniline point, net heat of combustion and boiling range were experimentally determined by the following method. Pensky-Martens closed cup ashpoint tester (SYP1002-II, Shanghai Petroleum Instrument Co. LD) was used to determine the ash point according to ASTM D93 method. Boiling point range was determined by boiling range analyzer (SYP2001-I, Shanghai

Table 1 Summary of hydrocarbon classes in jet fuels by GCMS Hydrocarbon classes Class Class Class Class Class 1 2 3 4 5 Descriptions Normal parans Isoparans Monocycloprans and alkylmonocycloprans Dicycloparans and alkyldicycloparans Hydroaromatics, including tetralins, alkyltetralins, decalin, alkyldecalin, etc. Benzene and alkylbenzenes Naphthalenes, alkylnaphthalenes and polycyclic aromatics Benzocycloparans, olens and other compounds

Class 6 Class 7 Class 8

Petroleum Instrument Co. LD) according to ASTM D86 method. Density, freezing point, aniline point and net heat of combustion were determined according to ASTM D 1298, ASTM D 2386, ASTM D611, ASTM D2382, respectively. Table A of in the supplement gives the composition property database of jet fuel used in this work. 3. Development of ANN correlation In this work, three types of ANN models were used as follows: single-layer feedforward neural network (SLFNN), multiple-layer feedforward neural network (MLFNN), and GRNN. All neural network calculations were performed using Matlab 6.5 software. As the rst step of ANN model development, the database was divided into training set and testing set to avoid the over-tting problem. The training set contained 58 data obtained with special carefulness to include the maximum and minimum value of input and target variables, and testing set contains 22 data to conrm the robustness of trained ANN. The selection strategy is that the 9 rows in table A consisting of the maximum and minimum values of input and target variables are rst found, and then randomly select 30% remaining data as testing set. More details about the training and testing data are indicated in Table A in the Supplementary data.

Abundance

1.3e+07 1.2e+07 1.1e+07 1e+07 9000000 8000000 7000000 6000000 5000000 4000000 3000000 2000000 1000000 5.00 10.00 15.00 20.00 25.00 30.00 35.00 40.00

Time-->

Fig. 1. Typical total ionization chromatography (TIC) of jet fuel sample.

2554

G. Liu et al. / Fuel 86 (2007) 25512559

C1 C2 P C8 T1 T2
Fig. 2. Architecture of single-layer feed-forward neural network (SLFNN) for jet fuel compositionproperty correlation.

The target variables were also normalized and used as output variables during the training and testing steps. The normalized output was converted into the desired value using the inverse function of normalization after property prediction. 4. Results and discussion In this work, compositionproperty of jet fuels were rstly correlated using MLR correlations to examine whether it was necessary or not to take the boiling ranges and isoparans content into the correlations. And then several ANN correlations developed above were used for the properties estimations from the jet fuels compositions. Moreover attempts were also made to compare prediction accuracies of MLR and ANN correlations, and thus to provide a set of reliable tools for the design and quality control of jet fuel. 4.1. MLR correlations In general, MLR correlations were used to correlate compositionproperty of fuels due to the concision of their equations [311]. In this work, attempts were rstly made to compare ve MLR correlations, in which three of them were proposed by Cookson et al. [3,4,7] and the other two equations developed in this work. Table 2 presents regression results of those MLR correlations. In general, for all the MLR correlations R2 reach up to only 0.9 in the best case. For our experimental data R2 of Cookson et al. [3,4,7] correlations are relatively lower than those reported in the original papers. This may be a result of the fact that inevitable deviations in the determination of the hydrocarbon class and concentration brought by GCMS and 13C NMR analysis methods. Eq. (4) yields higher R2 when compared with Eqs. (1) and (2), which further conrms that it is crucial to introduce the isoparans into the compositionproperty correlations. Incorporation of the isoparan and cycloalkane contents into Eq. (3) leads to Eq. (5) with higher R2, indicating that detailed hydrocarbon group is benecial for higher prediction accuracy. Cookson et al. [7] studied smoke point, freezing point, density, aromatics and hydrogen content of 72 jet fuels with the boiling range of 150250 C and core boiling range (T10 T90) of 190230 C, and reported that Eq. (3) was more eective than Eq. (1) for the low-temperature properties, such as freezing point. A similar result is also observed in this work. However, it is interesting to nd that for ash point predictions Eqs. (3) and (5) with core boiling range, as compared with Eqs. (1) and (4), can sharply increase regression R2 from 0.47 to 0.74, and form 0.51 to 0.74, respectively. Furthermore, for the prediction of the net heat of combustion, slight improvements are also observed in this research. From a practical point of view, it is important for the boiling range to be incorporated into the composition property relations for the ash point, low-temperature properties and net heat of combustion, etc.

Fig. 2 shows a schematic representation of SLFNN, which contains three layers, i.e., one input layer, one hidden layers and one output layer. The values for the hidden layer nodes of ANN were obtained by substituting the values of the normalized inputs and the corresponding connectivity weights into the input-hidden layer transfer function (logsig), and the normalized output was derived from hidden-output layer connectivity weights and the corresponding transfer function (tansig). In the training step, the SLFNN was supplied with a training set including the input and corresponding output values. The network correlated the inputs and outputs by nding the optimum set of weights to minimize the objective function. Dierent training algorithms available in the Matlab were tested. The traingdx algorithm was found to train the ANNs in a fast and robust way, avoiding the well-known over-tting problem. MLFNN used in this work was a four-layer backpropagation network consisting of two hidden layers. The transfer functions were logsig, tansig, and tansig. For both SLFNN and MLFNN, the network topology (i.e., the number of the input variables, the number and size of the hidden layers), was a critical parameter aecting the performance of the model. The topology of the developed ANNs were determined with repetitive trials, keeping the architectures as simple as possible to avoid overtraining due to the limited number of samples available in both training and test sets. The model with the smallest prediction error on the test data set after training was used to predict the test set and to verify the accuracy of the model. GRNN, a three-layer network with the hidden neuron number as many as the number of input samples, rst memorized all the patterns of the training set and then generated a group of smoothing factors representing the importance of each input to the model. When a test set was supplied to GRNN, smoothing factors groups was adjusted to nd new groups that best t to the test data set. Detailed descriptions of how neural networks work were presented elsewhere [23]. The jet fuel compositions in term of eight hydrocarbon groups, as well as the core boiling ranges (T10 and T90, i.e., temperature at 10% and 90% of distillates volume), were established as ANN input variables after the necessary pretreatment, such as normalization of T10 and T90.

G. Liu et al. / Fuel 86 (2007) 25512559 Table 2 Multiple linear correlations for compositionproperties of jet fuels Eqs. 1
a

2555

Properties Ta (C) D (g/cm3) T (C) Tfr (C) Hnc (MJ/kg) Ta (C) D (g/cm3) T (C) Tfr (C) Hnc (MJ/kg) Ta (C) D (g/cm3) T (C) Tfr (C) Hnc (MJ/kg) Ta (C) D (g/cm3) T (C) Tfr (C) Hnc (MJ/kg) Ta (C) D (g/cm3) T (C) Tfr (C) Hnc (MJ/kg)

Correlations 6.1Cn 33.45Car + 73.9 0.0561Cn 0.0075Car + 0.8307 2.0Cn 55.4Car + 0.47 77.58Cn + 17.07Car 73.76 0.63Cn 0.64Car + 43.31 8.66Cn + 3.03Cbc 30.79Car + 71.16 0.1242Cn 0.0810Cbc 0.0783Car + 0.9030 34.5Cn 38.6Cbc 89.2Car + 115.2 72.7Cn 5.8Cbc + 12.0Car 68.6 1.34Cn + 0.85Cbc + 0.10Car + 42.55 0.3Cn + 0.15Car + 8.45T10 55.11T90 + 101.8 0.0013Cn 0.0001Car 0.0454T10 + 0.0964T90 + 0.5593 0.91Cn 0.076Car + 4.1837T10 + 15.66T90 97.92 0.65Cn 0.25Car + 75.26T10 + 64.57T90 152.7 0.017Cn + 0.0039Car + 0.5977T10 1.9372T90 + 46.04 29.14Cn + 32.15 Ciso + 8.57Ccycl 13.61Car + 51.86 0.1713Cn 0.148Ciso 0.0937Ccycl 0.1178Car + 0.9473 46.14Cn 55.21Ciso 41.77Ccycl 98.96Car + 126.25 53.88Cn 32.53Ciso 10.91Ccycl 3.79Car 50.86 2.18Cn + 2.04Ciso + 1.07Ccycl + 0.80Car + 41.76 28.5Cn + 31.0Ciso + 8.6Ccycl 17.8Car 0.0498T10 + 0.0182T90 + 58.8 0.0614Cn 0.0403Ciso + 0.0193Ccycl + 0.0359Car + 0.0006T10 + 0.0003T90 + 0.6203 12.82Cn + 8.94Ciso + 11.09Ccycl + 23.01Car + 0.8836T10 0.0479T90 108.86 59.2Cn 21.14Ciso 13.04Ccycl + 41.99Car + 0.557T10 0.212T90 124.56 1.36Cn + 1.23Ciso + 0.25Ccycl 0.43Car 0.0059 T10 0.0021 T90 + 44.32

r2 0.6058 0.5791 0.4739 0.8673 0.4956 0.6063 0.6029 0.4906 0.8674 0.6150 0.7273 0.7822 0.7472 0.8939 0.6861 0.9083 0.7267 0.5140 0.8823 0.8238 0.9102 0.8766 0.7484 0.8958 0.8932

References [3]

2b

[4]

3c

[7]

4d

This work

5e

This work

a b c d e

P = a1Cn + a2Car + k0, wherein Car is equal to the sum of class 7 and 8. P = a1Cn + a2Cbc + a3Car + k0, wherein Cbc is equal to the sum of class 2, 3, 4 and 5. P = a1Cn + a2Car + k1T10 + k2T90 + k0, wherein T10 and T90 are core boiling range of 10% and 90%, respectively. P = a1Cn + a2Ciso + a3Ccycl + a4Car + k0, wherein Ccycl is equal to the sum of class 3, 4 and 5. P = a1Cn + a2Ciso + a3Ccycl + a4Car + k1T10 + k2T90 + k0.

4.2. ANN correlations 4.2.1. Density Density of jet fuel is dependent on its chemical composition. It is known that for a class of compounds the density increases with the carbon number (or boiling point range), while at the same carbon number, the density increases in the following order: parans < cycloparafns < aromatics. Therefore, both hydrocarbon type and boiling range are indispensable for the density estimation. Table 3 presents a statistical comparison of dierent compositiondensity relations consisting of the ANN models and the MLR Eqs. (4) and (5). It is evident that SLFNN, GRNN and MLFNN predict the densities of jet

fuels with lower mean of absolute relative errors (MARE) and standard deviations (SD) than the MLR Eqs. (4) and (5). For the whole database the SLFNN model is appropriate for the density prediction with MARE of 0.2052% and SD of 0.1738%. SLFNN predicts the density of jet fuel with the similar accuracy to the MLR correlation of Cookson et al. [3,4,7] and the ANN methods of Yang et al. [17,18]. Fig. 3 depicts a parity plot of the experimental versus predicting densities with SLFNN and the MLR Eqs. (4) and (5). It is indicated that the SLFNN network estimate more than 98% of the total data within a relative error envelope of 0.05%, while only (or less than) 90% of the data are predicted by the MLR Eqs. (4) and (5) falling within this region.

Table 3 Comparison of compositiondensity correlations for jet fuels Correlations SLFNNa GRNN MLFNNb Eq. (4) Eq. (5)
a b

Mean of absolute relative errors (%) Training set 0.1992 0.1764 0.1960 Testing set 0.2212 0.3886 0.2587 Total set 0.2052 0.2210 0.2133 0.3842 0.2634

Standard deviation (%) Training set 0.1821 0.1654 0.1754 Testing set 0.1524 0.2802 0.2227 Total set 0.1738 0.2143 0.2133 0.4398 0.2946

With 10 hidden neurons. Two hidden layers with 6 and 4 neurons.

2556
0.84 0.83 0.82 0.81 0.80 0.79 0.79 0.80 0.81 0.82
3

G. Liu et al. / Fuel 86 (2007) 25512559

90
SLFNN, R =0.9380 2 Equation (4), R =0.9083 2 Equation (5), R =0.9102 +0.05% -0.05%
2

Measured density, g/cm

Predicted flash point, C

85 80 75 70 65

0.83

0.84

MLFNN, R =0.9047 2 Equation (4), R =0.5140 2 Equation (5), R =0.7484

+3% -3%

65

70

75

80
o

85

90

Measured density, g/cm

Measured flash point, C

Fig. 4. Plot of experimental versus predicting ash point with MLFNN and MLR.

Fig. 3. Plot of experimental versus predicting density with SLFNN and MLR.

4.2.2. Flash point Flash point of jet fuel is the temperature at which the vapor pressure of the substance is such as to produce a concentration of vapor in the air that corresponds to the lower ammable limit, and thus is one of the major physical properties used to determine the ammability and volatilization of jet fuels. Caswell et al. [10] rst gave a ash point prediction method with LC/1H NMR but with R2 lower 0.72. Gomez-Carracedo et al. [12] suggested a ash point prediction method from FTIR spectra of vapour-phase generated by aviation jet fuel in the 4000600 cm1 wavenumber region, which can not be strictly considered as a composition-ash point relation. Therefore, it is still an important challenge to develop composition-ash point relation for jet fuel. Table 4 presents a statistical comparison of dierent composition-ash point relations with the ANN models and the MLR Eqs. (4) and (5). It is clear that SLFNN, GRNN and MLFNN predict the ash point of jet fuel with MARE and SD lower than the MLR Eqs. (4) and (5). For the whole database MLFNN is appropriate for the ash point prediction with MARE of 1.59% and SD of 01.52%. Fig. 4 depicts a parity plot of the experimental versus predicted ash point with MLFNN, as well as the MLR Eqs. (4) and (5). It is observed that MLFNN predict more than 90% of the data within the relative errors envelopes of 3%, while only (or less than) 6070% of the data
Table 4 Comparison of compositionash point correlations for jet fuels Correlations
a

predicted by the MLR Equations falling into this regime. MLFNN is also advantageous over Caswell et al. [10] correlation due to higher R2 of 0.90. Consequently, MLFNN is more reliable to predict the ash point of jet fuels. 4.2.3. Freezing point Freezing point is an essential parameter to show the lower temperature properties of jet fuel. Aens et al. [24] showed that the interactions of hydrocarbons in the fuel are so complex that there is sometimes substantial disagreement between literature and data derived Vant Ho equation. To all our knowledge, up to know only Cookson et al. [4,7,8] provided several MLR correlations for predicting freezing point with average R2 less than 0.8 using GC + HPLC, 13C NMR + HPLC and 1H NMR. It is worth doing more work and developing new compositionfreezing point correlation. Table 5 presents a statistical comparison of dierent compositionfreezing point relations consisting of the ANN models and the MLR Eqs. (4) and (5). It is obvious that SLFNN, GRNN and MLFNN correlations predict jet fuel freezing point with lower mean of absolute errors (MAE) and SD than the MLR Eqs. (4) and (5). Moreover, for the whole database the SLFNN model is appropriate for the freezing point prediction with MAE of 1.7 C and SD of 2.0 C. Fig. 5 depicts a parity plot of the experimental versus predicting freezing points with SLFNN and the

Mean of absolute relative errors (%) Training set Testing set 2.46 2.62 1.78 Total set 1.85 1.89 1.59 3.93 2.83

Standard deviation (%) Training set 1.37 1.41 1.50 Testing set 2.15 1.71 1.58 Total set 1.65 1.55 1.52 2.69 1.93

SLFNN GRNN MLFNNb Eq. (4) Eq. (5)

a b

1.62 1.61 1.53

With 11 hidden neurons. Two hidden layers with 8 and 4 neurons.

G. Liu et al. / Fuel 86 (2007) 25512559 Table 5 Comparison of compositionfreezing point correlations for jet fuels Correlations SLFNNa MLFNNb GRNN Eq. (4) Eq. (5)
a b

2557

Mean of absolute errors (C) Training set 1.5 1.9 1.7 Testing set 2.2 3.4 4.4 Total set 1.7 2.4 2.5 2.7 2.6

Standard deviation (C) Training set 1.9 1.8 1.4 Testing set 2.2 2.6 3.1 Total set 2.0 2.2 2.3 2.3 2.0

With 11 hidden neurons. Two hidden layers with 7 and 4 neurons.

-20

82 SLFNN, R =0.9051 2 Equation (4), R =0.8823 2 Equation (5), R =0.8758

o 2

80

Predicted freezing point, C

-30

Predicted aniline point, C

78 76

MLFNN, R =0.9513 2 Equation (4), R =0.9083 2 Equation (5), R =0.9102 +1.5% -1.5%

-40

+3.5 C -50 -3.5 C

o

74 72 70 68

-60

-70 -70

-60

-50

-40
o

-30

-20

68

70

72

74

76

78
o

80

82

Measured freezing point, C

Fig. 5. Plot of experimental versus predicting freezing point with SLFNN and MLR.

Measured aniline point, C

Fig. 6. Plot of experimental versus predicting aniline point with MLFNN and MLR.

MLR Eqs. (4) and (5). It is illustrated that more than 90% of the data estimated by the SLFNN fall within a deviation of 3.5 C, while only (or less than) 80% of the data predicted by the MLR Eqs. (4) and (5) fall into the same region. In general, this correlation predicts freezing point with R2 greater than 0.9, which is more reliable than the correlations of Cookson et al. [4,7,8]. 4.2.4. Aniline point Aniline point is the lowest temperature at which equal volumes of aniline and hydrocarbon sample are completely miscible. Aniline point determination is useful in characterizing pure hydrocarbons type in the jet fuel and estimating combustion heat.
Table 6 Comparison of compositionaniline point correlations for jet fuels Correlations
a

Table 6 presents performances of dierent composition aniline point relationships. Obviously, for the whole database the GRNN and MLFNN correlations are preferable compared with the MLR Equations due to lower MARE and SD. GRNN has a poor prediction capability in view of larger predicting deviation between the testing and training set. Fig. 6 depicts a parity plot of the experimental versus predicting the aniline point with MLFNN and the MLR Eqs. (4) and (5). It is shown that MLFNN predicts more than 95% of the data with a relative deviation of 1.5%, while only (or less than) 85% of the data predicted by the MLR equations falling within this envelope. Moreover, the MLFNN correlation predicts freezing point with higher accuracy (R2 > 0.95) than the correlation of Cookson et al. [5,7,8].

Mean of absolute relative errors (%) Training set Testing set 0.74 1.45 0.67 Total set 0.71 0.59 0.56 0.82 0.81

Standard deviation (%) Training set 0.56 0.28 0.44 Testing set 0.58 1.24 0.53 Total set 0.57 0.86 0.47 0.60 0.59

SLFNN GRNN MLFNNb Eq. (4) Eq. (5)

a b

0.70 0.27 0.57

With 10 hidden neurons. Two hidden layers with 6 and 4 neurons.

2558

G. Liu et al. / Fuel 86 (2007) 25512559

Table 7 Comparison of compositionnet heat of combustion correlations for jet fuels Correlations SLFNNa GRNN MLFNNb Eq. (4) Eq. (5)
a b

Mean of absolute errors (102 MJ/kg) Training set 2.0 1.9 2.3 Testing set 2.7 4.4 2.4 Total set 2.2 2.6 2.3 3.2 2.6

Standard deviation (102 MJ/kg) Training set 1.7 1.7 1.9 Testing set 2.1 3.4 1.4 Total set 1.8 2.6 1.8 2.8 2.1

With 11 hidden neurons. Two hidden layers with 7 and 4 neurons.

Predicted net heat of combustion, MJ/kg

43.8 43.7 43.6 43.5 43.4 43.3 43.2 43.2

+0.045 MJ/kg

SLFNN, R =0.9278 2 Equation (4), R =0.8238 2 Equation (5), R =0.8932

Table 8 Repeatability of ASTM and ANN methodsa Properties D (g/cm ) T (C) Tfr (C) Ta (C) Hnc (MJ/kg)
a b 3

ASTMb 0.0005 0.6 0.9 0.7 0.05

ANN 0.0028 0.7 0.9 0.4 0.03

-0.045 MJ/kg

Repeatability is dened as the standard deviation of 5 tests. ASTM method for the density, freezing point, ash point, aniline point and net heat of combustion are D 1298, D 2386, D93, D611, D2382, respectively.
43.3 43.4 43.5 43.6 43.7 43.8

Measured net heat of combustion, MJ/kg

Fig. 7. Plot of experimental versus predicting net heat of combustion with SLFNN and MLR.

4.2.5. Net heat of combustion Net heat of combustion is one of important specication properties of jet fuel accounting for the combustion performance. Cookson et al. [4,5,8], Caswell et al. [10] and Fodor and Kohl [11] suggested four correlations to estimate the net heat of combustion from dierent techniques. Table 7 presents performances of the ANN and MLR approaches on chemical compositionnet heat of combustion relationships. Obviously, SL- and MLFNN correlations are appropriate in view of their lower MARE and SD, as compared with GRNN. For the whole database SLFNN model is an ideal model for the prediction of the net heat of combustion with MAE of 0.022 MJ/kg and SD of 0.018 MJ/kg. Fig. 7 gives a parity plot of the experimental versus predicting net heat of combustion with the SLFNN and MLR Eqs. (4) and (5). It is indicated that more than 91% of the data estimated by the SLFNN fall into the MRE envelope of 0.045 MJ/kg, while for the MLR equations only (or less than) 80% of the data. The SLFNN correlation predicts the net heat of combustion of jet fuel with (R2 > 0.92), an accuracy similar to the correlations of Caswell et al. [10] and Fodor and Kohl [11], and higher than Cookson et al. [4,5,8]. 4.3. Comparison of ASTM and ANN methods In Table 8, the repeatability of the recommended ANNs and the ASTM test methods are presented. It is clear that

the repeatability of neural network models for density and ash point are lower than the ASTM test methods. However, for the freezing point, aniline point and net heat of combustion the repeatabilities of the ANN methods are as good as the ASTM methods. Comparing the repeatability of the ANNs with the ASTM methods, it can be concluded that the ANN approaches based on GCMS are comparable to the repeatability values of the standard ASTM methods, which are used for the experimental determination of the properties. As shown by Yang et al. [18], a larger database and a more careful identication of hydrocarbon groups will lead to a substantial improvement of the neural network correlation, and the ANN correlations could probably be further improved by introducing more inputs or slightly dierent inputs to the correlations. Therefore, the developed ANNs are eective in predicting the jet fuel properties, which provides a rapid alternative to the laborious standard methods. 5. Conclusions A jet fuel compositionproperty relation database consisting of 80 samples with the corresponding properties (aniline point, density, freezing point, ashpoint, and net heat of combustion) was developed using GCMS to analyze and group fuel compositions into eight hydrocarbon compound classes, including normal parans, isoparans, monocycloprans, dicycloprans, alkylbenzens, naphthalenes, tetralins, hydroaromatics. Several MLR equations reported in the literatures and proposed by the authors were rstly used to correlate the

G. Liu et al. / Fuel 86 (2007) 25512559

2559

compositionproperty relations. It was interesting to nd that the isoparans concentrations and boiling ranges are two essential factors for improving MLR prediction accuracy of some properties, such as ash pinot, aniline point and freezing point. Novel and more general ANN correlations were developed with SLFNN, GRNN and MLFNN models. Eight mass fractions of hydrocarbon groups and two normalized temperatures of core boiling range were used as the ANN inputs. The performance comparisons of ANN and MLR methods indicated that ANN are more general methods with signicant improvements on MLR equations in correlating jet fuel compositionproperty relationships. It was also concluded that SLFNNs are preferable to predict of density, freezing point and net heat of combustion, while MLFNNs are appropriate for ash point and aniline point prediction. Comparing the repeatability of the ANNs with the ASTM methods, it was concluded that the ANN approaches based on GCMS are comparable to the repeatability values of the standard ASTM methods. The developed ANN approaches on compositionproperty relationships is a reliable and promising tool for the development of next generation jet fuel and the rapid quality control technique with higher reliability. Acknowledgements Financial supported by National Hi-Tech Research and Development 863 Program of China (No.2005AA723101), the Programme of Introducing Talents of Discipline to Universities (No. B06006), China Postdoctoral Science Foundation (20060400191) and Innovative Research Project of Petrochina Co. Ltd are gratefully acknowledged. In addition, GL want to express his special thanks to Professor Akira Tomita and the reviewers for their helpful academic suggestions and to Ms. Sini of Tianjin University for her careful work in polishing English of this MS. Supplementary data Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j.fuel.2007. 02.023. References
[1] Maurice LQ, Lander H, Edwards T, Harrison WE. Advanced aviation fuels: a look ahead via a historical perspective. Fuel 2001;80:74756. [2] Edwards T. Liquid fuels and propellants for aerospace propulsion: 19032003. J Propul Power 2003;19:1089107. [3] Cookson DJ, Latten JL, Shaw IM, Smith BE. Propertycomposition relationships for diesel and kerosene fuels. Fuel 1985;64:50919.

[4] Cookson DJ, Lloyd CP, Smith BE. Investigation of the chemical basis of kerosene (jet fuel) specication properties. Energy Fuel 1987;1:43847. [5] Cookson DJ, Lloyd CP, Smith BE. Investigation of the chemical basis of diesel fuel properties. Energy Fuel 1988;2:85460. [6] Cookson DJ, Smith BE, Johnston RRM. Relationships between diesel fuel ignition quality indicators and composition. Fuel 1993;72:6614. [7] Cookson DJ, Iliopoulos P, Smith BE. Compositionproperty relations for jet and diesel fuels of variable boiling range. Fuel 1995;74:708. [8] Cookson DJ, Smith BE. Calculation of jet and diesel fuel properties using 13C NMR spectroscopy. Energy Fuel 1990;4:1526. [9] Cookson DJ, Smith BE. Observed and predicted properties of jet and diesel fuels formulated from coal liquefaction and FischerTropsch feedstocks. Energy Fuel 1992;6:5815. [10] Caswell KA, Glass TE, Swann M, Dorn HC. Rapid prediction of various physical properties for middle distillate fuels utilizing directly coupled liquid chromatography/1H nuclear magnetic resonance. Anal Chem 1989;61:20611. [11] Fodor GE, Kohl KB. Analysis of middle distillate fuels by midband infrared spectroscopy. Energy Fuel 1993;7:598601. [12] Gomez-Carracedo MP, Andrade JM, Calvino M, Fernandez E, Prada D, Muniategui S. Multivariate prediction of eight kerosene properties employing vapour-phase mid-infrared spectrometry. Fuel 2003;82:12118. [13] Pasadakisa N, Sourligasa S, Foteinopoulos C. Prediction of the distillation prole and cold properties of diesel fuels using mid-IR spectroscopy and neural networks. Fuel 2006;85:11317. [14] Karonis D, Lois E, Zannikos F, Alexandridis A, Sarimveis H. A neural network approach for the correlation of exhaust emissions from a diesel engine with diesel fuel properties. Energy Fuel 2003;17:125965. [15] Liu G, Mi Z. Prediction of pulsation frequency of pulsing ow in trickle-bed reactors using articial neural network. Chem Eng Sci 2004;59:578794. [16] Ramadhas AS, Jayaraj S, Muraleedharan C, Padmakumari K. Articial neural networks used for the prediction of the cetane number of biodiesel. Renew Energy 2006;31:252433. [17] Yang H, Ring Z, Briker Y, McLean N, Friesen W, Fairbridge C. Neural network prediction of cetane number and density of diesel fuel from its chemical composition determined by LC and GCMS. Fuel 2002;81:6574. [18] Yang H, Briker Y, Szynkarczuk R, Ring Z. Prediction of density and cetane number of diesel fuel from GC-FIMS and PIONA hydrocarbon composition by neural network. Prepr Pap Am Chem Soc Div Fuel Chem 2004;49:813. [19] Briker Y, Ring Z, Iacchelli A, McLean N, Rahimi PM, Fairbridge C, et al. Diesel fuel analysis by GC-FIMS: aromatics, n-parans, and isoparans. Energy Fuel 2001;15:2337. [20] Briker Y, Ring Z, Iacchelli A, McLean N, Fairbridge C, Malhotra R, et al. Diesel fuel analysis by GC-FIMS: normal parans, isoparafns, and cycloparans. Energy Fuel 2001;15:9961002. [21] Katherine DD, Campbell JL, Bartlett MG, Witten ML, Fisher JW. Validation of a gas chromatography/mass spectrometry method for the quantication of aerosolized jet propellant. J Chromatogr A 2005;1093:1120. [22] Yu J, Eser S. Thermal decomposition of jet fuel model compounds under near-critical and supercritical conditions. 2. Decalin and tetralin. Ind Eng Chem Res 1998;37:46018. [23] Haykin S. Neural networks a comprehensive foundation. New Delhi: Prentice-Hall Inc; 1999. [24] Aens WA, Hall JM, Holt S, Hazlett RN. Eect of composition on freezing points of model hydrocarbon fuels. Fuel 1984;63:5437.