Professional Documents
Culture Documents
Fuel
journal homepage: www.elsevier.com/locate/fuel
Prediction of HEFA content in jet fuel using FTIR and chemometric methods T
b,⁎
a
Dan Vrtiška , Petr Vozka , Veronika Váchová , Pavel Šimáček , Gozdem Kilaz
a a b
a
Department of Petroleum Technology and Alternative Fuels, University of Chemistry and Technology, Prague, Technicka 5, Prague 6 16628, Czech Republic
b
School of Engineering Technology, Fuel Laboratory of Renewable Energy (FLORE), Purdue University, 401 N Grant St., West Lafayette 47907, IN, USA
G R A P H I C A L A B S T R A C T
A R T I C LE I N FO A B S T R A C T
Keywords: Alternative aviation fuel blending components are becoming a common component in gas turbine engines fuels.
Biojet One of the blending components is Hydroprocessed Esters and Fatty Acids (HEFA). Predictive models based on a
HEFA chemometric processing of FTIR spectra were developed as a reliable method for the determination of HEFA in
Jet fuel petroleum jet fuel/HEFA blends. The concentration of HEFA in blends with jet fuel ranged between 0 and 60 wt
FTIR spectroscopy
%. The sample set composed of calibration and validation subsets. Additionally, several pre-processing techni-
Principal component analysis
ques were evaluated, such as mean centering, variance scaling, derivation, and smoothing. The effectiveness of
Partial least square regression
these models was highly dependent on the model parameters (pre-processing techniques and latent variables).
The most robust model was able to predict HEFA content in all samples with a good predictive ability (root mean
square error of prediction of 0.7 wt%).
⁎
Corresponding author.
E-mail addresses: vrtiskad@vscht.cz (D. Vrtiška), pvozka@purdue.edu (P. Vozka), vachovav@vscht.cz (V. Váchová), simacekp@vscht.cz (P. Šimáček),
gkilaz@purdue.edu (G. Kilaz).
https://doi.org/10.1016/j.fuel.2018.09.102
Received 27 June 2018; Received in revised form 7 September 2018; Accepted 22 September 2018
Available online 29 September 2018
0016-2361/ © 2018 Elsevier Ltd. All rights reserved.
D. Vrtiška et al. Fuel 236 (2019) 1458–1464
of this writing, five alternative aviation blending components have been Table 1
approved in varying ratios: Fischer-Tropsch Hydroprocessed Synthe- Fuel samples together with their origin and manufacturer.
sized Paraffinic Kerosine (FT-SPK), Synthesized Paraffinic Kerosine Sample Fuel Origin Manufacturer/Supplier
from Hydroprocessed Esters and Fatty Acids (HEFA), Synthesized Iso-
Paraffins from Hydroprocessed Fermented Sugars (SIP), Synthesized J1 Jet A-1 petroleum-based Unipetrol, a.s., Czech Republic
J2 Jet A-1 petroleum-based Twin Trans, s.r.o, Czech Republic
Kerosine with Aromatics Derived by Alkylation of Light Aromatics from
J3 Jet A petroleum-based Shell, USA
Non-Petroleum Sources (SPK/A), and Alcohol-to-Jet Synthetic Paraf- J4 Jet A petroleum-based Exxon Mobil, USA
finic Kerosene (ATJ) [4]. None of these components, except SPK/A in H1 HEFA camelina UOP, USA
Europe, can be utilized as neat (100%) drop-in replacement for jet fuel H2 HEFA tallow UOP, USA
in current gas turbine engines due to their property limitations. These H3 HEFA mixed fats Dynamic Fuels, USA
1459
D. Vrtiška et al. Fuel 236 (2019) 1458–1464
Table 2
Properties of neat jet fuel samples.
Parameter Method Neat jet fuel samples
J1 J2 J3 J4 H1 H2 H3
a
n-Paraffins (wt%) GC × GC-FID 23.4 22.8 21.5 21.0 8.5 8.9 10.1
iso-Paraffins (wt%) GC × GC-FIDa 27.5 32.1 33.1 30.6 89.7 90.5 88.5
Cycloparaffins (wt %) GC × GC-FIDa 28.4 27.5 27.6 24.4 1.7 0.6 1.4
Aromatics (wt%) GC × GC-FIDa 20.7 17.6 17.8 24.0 0.03 0.01 0.02
Hydrogen content (wt%) ASTM D7171 14.0 14.2 14.1 13.8 15.4 15.4 15.1
Density at 15 °C (kg·m−3) ASTM D4052 802.5 792.9 804.0 805.7 759.8 757.3 761.2
Kin. viscosity at −20 °C (mm2·s−1) ASTM D7042 3.8 3.2 4.7 3.8 5.2 5.0 5.6
Freezing point (°C) ASTM D2386 −54 −63 −50 −51 −55 −59 −49
Flash point (°C) ASTM D56 50 42 48 43.0 42.5 52.5 41.5
Net heat of combustion (MJ·kg−1) ASTM D4809 43.3 43.3 43.3 43.1 44.0 44.2 44.2
Distillation temperature: ASTM D86
10 vol.% recovery (°C) 179 170 176 170 161 173 179
50 vol.% recovery (°C) 196 186 204 197 214 210 222
90 vol.% recovery (°C) 219 216 245 248 269 242 257
Final boiling point (°C) 233 243 269 266 281 254 273
a
GC × GC-FID parameters can be found in [16].
X (J1-H1) Y (J2-H1) Z (J3-H1) this method is also composed of data compression to the new variables,
these variables, called factors or latent variables, are calculated dif-
C (J4-H1) H (J4-H2) R (J4-H3) ferently than in the PCA. Unlike PCA, PLS is working with two data
matrices (e.g. spectral data and concentrations). The latent variables
are obtained by maximizing the covariance between both matrices,
which means that these variables are more closely related to the matrix
with concentrations than in the regression method based on PCs, which
is called principal components regression [18–21].
In addition to the two above-mentioned chemometric methods,
several techniques of spectra preprocessing (mean centering, variance
scaling, derivation, and smoothing) were also employed and an optimal
combination was sought.
Mean centering is a preprocessing tool that subtracts the average
from each variable (e.g. absorbance). This operation modified relative
differences in the value of each affected variable so a better distinction
of the spectra can be observed [18,19]. When variance scaling is uti-
lized, each variable is divided by the value of its standard deviation.
Since the scale of the variables is unified, even the variables with the
low signal intensity achieve similar significance such as the variables
with the high intensity of their signal [18,19]. Derivation of the spectra
and using of their first or second derivative is commonly utilized pre-
processing tool for spectral scatter reduction which uncovers or high-
0 10 20 30 40 50 60 light feebly detectable and overlapped peaks. However, since spectral
HEFA content (wt. %) derivation usually decreases the overall signal intensity and S/N ratio,
due to the nature of derivatives calculation, it is highly encouraged to
Fig. 1. Overview of the subset compositions. include data smoothing method [18,20,21]. Savitzky-Golay (S-G)
smoothing technique is an effective method used for infrared spectra
investigate the relationship between neat jet fuel samples and the preprocessing [22]. During the S-G smoothing, the data within the
blends and partial least square regression (PLS) was employed to de- smoothed frame are interlaced by a suitable polynomial function at the
velop models for the quantification of HEFA in blends. bases of the least squares method. A number of S-G smoothing para-
PCA is an exploratory data analysis with the main goal of variable meters combinations (i.e., second, third, fourth, fifth, polynomial order
number reduction or data compression. For each data matrix composed and width of the smoothed frame) allows obtaining numerous models.
of m rows (samples) and n columns (variables), each row vector (e.g. A separate data smoothing technique, a method developed by Karl H.
FTIR spectrum) can be plotted in the n-coordinate system as a single Norris, is called Gap-Segment (G-S) derivative [23]. In the course of G-S
point. PCA replaces the original coordinate system by one composed of derivative, averages of variables within two segments (width is given by
principal components (PC). The direction of the first new variable (PC1) a number of variables) are acquired. When the difference between these
is adjusted so that it captures the largest part of the data variability. two averages is utilized for spectra derivation, a less noisy curve is
Since PCA aims to capture the highest data variability with the lowest obtained. Utilization of this smoothing technique offers many para-
number of the variables possible, every other PC captures the largest meters combinations such as the width of the smoothed segment and
part of the remaining variability. All PCs are obtained as linear com- the size of the space of segments of the smoothed points.
binations of the original variables and are mutually orthogonal (linearly Except for the spectral data preprocessing techniques briefly de-
uncorrelated). Every sample can then be plotted in the new coordinate scribed above, the number of latent variables used for the PLS model
system formed by PCs. The position of the samples in the new co- development is another important parameter of PLS models [24]. The
ordinate system is given by their score value [17–20]. most widespread procedure used to generate the optimal number of
PLS is a widely used method of multivariate calibration. Although latent variables is called cross-validation (CV); however, utilization of
1460
D. Vrtiška et al. Fuel 236 (2019) 1458–1464
the external validation procedure provides more meaningful results, for 1.5
example as demonstrated in [15,19]. The validation procedure based H-J blend
on the CV uses just the calibration standards, an independent validation JET J1
N
∑i = 1 (yp, i −ym, i )2
RMSE =
N (1)
J2
where N is the number of observations (number of samples), yp,i is 0.9
predicted the value of i-th observation and ym,i is a measured value of i- -9.5 -8.6 -7.7 -6.8 -5.9
th observation. The average difference between the actual and pre-
dicted values of the property of interest is given by the two RMSE PC1 score (89.5 % of variance)
parameters whose units are the same with the property being studied. Fig. 3. PCA score plot (PC1 vs. PC2) of spectral data.
3. Results and discussion attributed mainly to stretching of eC]Ce (Fig. 2a) and both in-plane
and out-of-plane deformation vibrations of C]CeH groups (Fig. 2b),
3.1. FTIR spectra which occur in the ring structure of aromatic hydrocarbons [25]. Ske-
letal vibrations of CeC groups and deformation vibrations of CeH
The recorded FTIR spectra of neat jet fuel samples (4 petroleum- groups can be observed in Fig. 2b for petroleum-derived fuels as well as
derived jet fuels and 3 HEFA jet fuels) are shown in Fig. 2. It is shown for HEFA samples.
that the spectra of the HEFA jet fuels (−) are different from those of the
petroleum-derived jet fuels (−). Although the FTIR spectra of the
samples were measured in the range of 4000–650 cm−1, only two 3.2. Principal component analysis (PCA)
narrower regions (a: 1633–1579 cm−1 and b: 1348–678 cm−1) were
used for further development of the models. The rest of the spectral The spectra of all samples utilized in this work served as input data
data was considered as irrelevant and was removed, especially due to for the PCA. As mentioned above, only two narrower spectral regions
the too strong absorbance caused by CeH stretching vibrations in the were used for the further processing. No preprocessing techniques were
region of 3000–2800 cm−1 and CeH deformation vibrations in the re- applied to the spectral data before PCA. Score plot of the first two
gion of 1480–1370 cm−1. Petroleum-derived jet fuels spectra contained principal components is shown in Fig. 3. These two principal compo-
more bands than the spectra of HEFA jet fuel. These bands can be nents captured 95.3% of spectra variance. All three samples of HEFA
1.0
a b
0.8
Absorbance (a.u.)
0.6
0.4
0.2
1633 1620 1606 1593 1579 1348 1214 1080 946 812 678
Wavenumber (cm-1) Wavenumber (cm-1)
Fig. 2. Two selected spectral regions used for the chemometric processing; spectra of HEFA jet fuels (−) and petroleum-derived jet fuels (−).
1461
D. Vrtiška et al. Fuel 236 (2019) 1458–1464
were located very close to each other on the left side of the plot. This Table 4
was simply an indication of their similarity. On the other hand, the Models parameters and results of the PLS models development.
samples of the petroleum-derived jet fuels were scattered on the right Model PLS-1 PLS-2 PLS-3 PLS-4 PLS-5
side of the plot. It means that the first principal component captured
sufficient spectral information to create separate clusters of HEFA fuel Number of latent variables 10 13 10 11 10
Preprocessing technique MCa – – – –
and petroleum-derived jet fuel which is in compliance with the results
Spectrum formatb S 2D 2D 2D 2D
obtained in the previous study where PCA of HVO fuel and diesel fuel Method of smoothingc,d S-G S-G S-G G-S G-S
spectra was presented [15]. It should be noted here that the distances Parameters of smoothingc,d 17–4 17–4 21–4 5–0 1–5
between petroleum-derived jet fuel sample points were higher than RMSECV (wt%) 0.08 0.21 0.17 0.15 0.16
those of HEFA. This indicated that petroleum-derived jet fuel chemical RMSEP (wt%) 0.03 0.18 0.54 0.12 0.70
1462
D. Vrtiška et al. Fuel 236 (2019) 1458–1464
0.14 1.2
Calibration Validation Calibration Validation
0 0
-0.07 -0.6
a c
-0.14 -1.2
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Actual HEFA content (wt. %) Actual HEFA content (wt. %)
0.5 0.4
Calibration Validation b d Calibration Validation
Calculation error (wt. %)
0 0
-0.25 -0.2
-0.5 -0.4
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Actual HEFA content (wt. %) Actual HEFA content (wt. %)
Fig. 4. Prediction errors between the calculated and actual HEFA content for models PLS-1 (a), PLS-2 (b), PLS-3 (c) and PLS-4 (d).
distribution of the calibration samples, PLS-3 model validation samples those of the petroleum-derived jet fuel samples. Partial least squares
prediction errors showed a different trend. The prediction errors in- regression was utilized in order to develop predictive models for HEFA
creased with increasing HEFA content in the samples, except for two content quantification in HEFA/jet fuel blends. The model that was
validation samples with the lowest HEFA content. The highest error was calibrated and validated using samples blended from the same neat jet
observed for the validation sample with the highest HEFA content, fuel samples showed very good predictive ability and negligible de-
which led to the high RMSEP value of this model. pendence on the preprocessing technique. RMSEP of this model was
The PLS-5 model was developed via simultaneously utilizing all 0.03 wt%. Predictive ability of the models validated by samples in-
three subsets used for the validation of PLS-2, PLS-3, and PLS-4 models. dependent on the calibration samples was satisfactory as well.
The lowest RMSEP value of PLS-5 was obtained using second derivative Nevertheless, the effectiveness of these models was highly dependent
spectra smoothed by a Gap-Segment method. The model parameters on the model parameters. When the models were validated by different
and results are shown in Table 4. Although the RMSEP value was higher sample subsets, each of those required unique combination of model
than those of PLS-2, PLS-3, and PLS-4 models, the robustness of this parameters. Last model (PLS-5) was developed in order to create a ro-
model was higher as, unlike the previous models, this model was able to bust model, whose predictive ability will be less dependent on model
predict HEFA content in all three validation subsets at once and with parameters. Although RMSEP of this model was higher (0.7 wt%) than
good predictive ability. Distribution of prediction errors within the the preceding ones, the model PLS-5 has a good predictive ability of the
entire range of calibration and validation samples is shown in Fig. 5. As HEFA content in all samples. This indicated the high robustness of this
expected, prediction errors of the validation samples were higher than model.
those of the calibration samples. Although the model is able to predict Although the multivariate calibration of spectral data, either NIR or
HEFA content in HEFA/jet fuel blends accurately, the variability of the MIR, can be utilized for the quantification of HEFA content in blends
validation sample chemical composition (mainly caused by variability with petroleum-based jet fuel, it is necessary to take into consideration
of petroleum-derived jet fuel samples) is responsible for the lower ef- the inconsistency of chemical composition of both fuels. The incon-
fectiveness of the model to predict HEFA content in validation samples sistency is responsible for the variability of the models prediction effi-
compare to calibration samples. ciency. Future steps in mitigating the challenges of HEFA content pre-
diction in jet fuels via FTIR and chemometrics will focus on the
extension of the calibration set by other independent samples. This is
4. Conclusion predicted to reduce the dependence on model parameters, which would
enable increase practical applications of this method.
In this study, chemometric methods were utilized to develop a
method for prediction of Hydroprocessed Esters and Fatty Acids (HEFA)
content in HEFA/petroleum-based jet fuel blends. The principal com- Acknowledgements
ponent analysis was used for investigating the variability in sample
chemical compositions of HEFA and petroleum-derived jet fuels. This work was supported by the National Program of Sustainability
Although the HEFA samples utilized in this study were produced from a (NPU I LO1613, MSMT-43760/2015) and specific university research
different feedstock, their chemical composition was less complex than (MSMT No 21-SVV/2018). Our special thanks extended to Dr. James
1463
D. Vrtiška et al. Fuel 236 (2019) 1458–1464
1.6
Calibration Validation
0.8
Calculation error (wt. %)
-0.8
-1.6
0 10 20 30 40 50 60
Actual HEFA content (wt. %)
Fig. 5. Prediction errors between the calculated and actual HEFA content within the entire range of calibration and validation samples (model PLS-5).
1464