jaap 2017 826 revision 1 v0 - consejo superior de
TRANSCRIPT
![Page 1: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/1.jpg)
Highlights
• Elemental analysis, water content and heating value were modeled.• Predictions were checked to prove the robustness of the models• Good predictions were obtained for all properties with R2
Pre≥0.82• Differences between predictions and experimental results were not statistically
significant
![Page 2: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/2.jpg)
![Page 3: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/3.jpg)
1
1 Prediction of elemental composition, water content and heating value of upgraded
2 biofuel from the catalytic cracking of pyrolysis bio-oil vapors by infrared
3 spectroscopy and partial least square regression models.
4 A. Veses, J.M. López, T. García, M.S. Callén*
5 Instituto de Carboquímica (ICB-CSIC), C/ Miguel Luesma Castán, 50018 Zaragoza,
6 Spain.
7 Abstract
8 The elemental composition, heating value and water content, are important properties to
9 be characterized for pyrolysis bio-oils, providing information on their quality. These
10 properties are mainly determined according to ASTM standards by using three different
11 analytical techniques requiring time and cost. This research was focused on a simple
12 method to determine the weight content of carbon, hydrogen, oxygen and water as well
13 as the heating value, by Fourier transform infrared spectroscopy (FT-IR) using models
14 based on partial least squares regressions (PLS). Samples were classified into two sets
15 according to Kennard-Stone algorithm. The first set of samples was used to develop the
16 calibration models for each physical parameter, where the number of latent variables
17 was determined by full cross validation procedure. The second set of samples was
18 employed as an external prediction set, assessing the quality of the models. External
19 predictions confirmed that robust models were developed since elemental analysis,
20 heating value and water content of the upgraded biofuels obtained by the catalytic
21 cracking of pyrolysis bio-oil vapors could be determined with good predictive ability
22 with a root mean square error of prediction of carbon content =0.963 wt.% (R2=0.836,
23 range=70.9-78.8 wt.%), hydrogen content = 0.101 wt.% (R2=0.815, range=8.01-8.75
24 wt.%), oxygen content=0.910 wt.% (R2=0.873, range=12.3-20.9 wt.%), water
25 content=0.416 wt.% (R2=0.829, range= 2.79-7.24 wt.%) and heating value=0.539 MJ
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859
![Page 4: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/4.jpg)
2
26 kg-1(R2=0.874, range= 31.9-36.9 MJ kg-1) by chemometric tools joined to medium
27 infrared spectrum.
28 Keywords: upgraded biofuels; infrared spectroscopy; elemental composition; heating
29 value; water content; partial least square regressions.
30 Corresponding author: Mª Soledad Callén. E-mail: [email protected]. Phone number:
31 +34 976 733977, Fax number: +34 976733318
32
33 1. Introduction
34 The applications of biomass for power and heat generation are becoming of increasing
35 interest due to its renewable character and the CO2 neutral overall balance in the
36 atmosphere [1]. There are different classes of biomass for biofuel production. Whilst
37 sugars and vegetable edible oils are already implemented to produce the so-called “first
38 generation” biofuels, lignocellulosic pyrolysis bio-oils are postulated for the production
39 of the “second generation” biofuels. It is generally accepted that first generation
40 biofuels have some important disadvantages not only related to the food versus fuel
41 debate [2], but also to their negative environmental impact and carbon balance [3].
42 These issues convert the second generation biofuels as potential alternatives.
43 In general, pyrolysis bio-oils can be considered as a complex mixture of chemicals
44 containing varied functional groups [4, 5] with numerous undesirable properties for fuel
45 applications like high water content, low stability, and remarkable acidity [6, 7]. Since
46 these bio-oils cannot be directly employed in the present infrastructures for power and
47 heat generation, bio-oils need to be upgraded by using catalytic processes, which
48 enhance the bio-oil quality by reducing the oxygen content as well as improving other
49 biofuel properties. As for other conventional fuels, their physical and chemical
50 properties provide information regarding the biofuel quality, stability and potential uses
60616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118
![Page 5: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/5.jpg)
3
51 as alternative fuels in conventional infrastructures [8]. These properties include water
52 content, pH, total acid number (TAN), heating value, elemental composition, density
53 and viscosity, among others.
54 Fuel properties are commonly measured by standards procedures ensuring quality and
55 comparability of results. These measurements require time and costly equipment to
56 perform the different analyses. One of the most important characteristics of upgraded
57 biofuels is the heating or calorific value, which is mainly measured by a bomb
58 calorimeter requiring a tedious, time-consuming process and advanced technical skills
59 in the handling of equipment [9-11]. The heating value can also be calculated by an
60 empirical model requiring the elemental analysis, in particular, the carbon, hydrogen
61 and oxygen content [12-14]. Precisely, the elemental analysis is one of the main
62 properties characterizing biofuels. In general, the procedure uses an analyzer which is
63 based on the flash combustion of dried and ground samples in combination with
64 chromatographic methods and a thermal conductivity detector (TCD) [15]. Other
65 important biofuel property is related to the water or moisture content, which requires its
66 determination with Karl Fischer titration [16-17]. This is a chemical analysis procedure
67 based on the oxidation of sulfur dioxide by iodine in a methanolic hydroxide solution.
68 This titration can be performed volumetrically or coulometrically, requiring an adaption
69 of the working method to the specific sample. Therefore, there is a request for general
70 analytical tools that allow predicting sample properties in a faster and easier way than
71 conventional techniques. One of these predicting tools is based on infrared spectroscopy
72 (IR) [18-21] combined with chemometric techniques [22-24]. Basically, IR needs trace
73 sample and it is a non-destructive analysis with minimal sample preparation, providing
74 fast and simultaneous analysis of multiple components from the same sample in a single
75 instrumental analysis. These advantages allow IR joint to chemometric tools to be
119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177
![Page 6: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/6.jpg)
4
76 successfully employed for control and monitoring operations in several sectors such as
77 agricultural, food and pharmaceutical industries [25, 26]. IR spectroscopy and
78 multivariate analysis have also shown to be an effective, rapid and accurate method for
79 predicting the composition of different biomass and biodiesel samples, as recently
80 shown in a comprehensive review by Chadwick et al. [27]. Additionally, heating value,
81 moisture content and ash content are other important parameters that have been
82 successfully determined for biomass quality [28]. The determination of hydrogen
83 content and heating value in diesel fuels has effectively been accomplished by Al-
84 Ghouti et al. [29] using Fourier transform infrared spectroscopy (FTIR) and PLS
85 models. IR coupled to PLS have also been applied to determine the total acid number
86 (TAN) in mineral gas engine oils [30]. The results from these studies have shown that
87 IR combined to chemometric tools may indeed be used to predict all of the above
88 properties at laboratory scale and for industrial samples, with the potential for online
89 monitoring and quality control purposes for the final product. However, there is not any
90 attempt to develop chemometric models to predict the main properties of pyrolysis
91 lignocellulosic biofuels in the literature.
92 Herein, the use of FTIR as a fast tool to predict different properties of biofuels obtained
93 from the catalytic upgrading of pyrolysis lignocellulosic bio-oils has been investigated
94 for the first time. Water content, elemental analysis (%C, %H and %O) and heating
95 value were predicted using chemometric tools based on linear regression models such as
96 partial least squares regressions, providing fast and cost-effective methods that reduce
97 time and expenses when compared to traditional methodologies.
98 2. Materials and methods
99 2.1. Upgraded bio-oils
178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236
![Page 7: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/7.jpg)
5
100 The biomass used in this work was forest pine woodchips (Pinus halepensis), including
101 bark, obtained from the North-East area of Spain. Raw bio-oil used in the catalytic
102 upgrading tests was obtained from the catalytic pyrolysis of biomass in an auger reactor
103 operated at 450ºC with N2 as the carrier gas. The residence time of solids in the reactor
104 was 7 min. Tests were carried out with a catalyst (90% CaO, Calcinor) to biomass mass
105 ratio of 1:6 and catalysts were diluted with sand, keeping a (sand + catalyst) to biomass
106 mass ratio of 3:1. The liquid obtained was separated in two different layers by
107 centrifugation at 1500 rpm for 1 hour: the upper or aqueous layer and the bottom or
108 organic layer, which was selected as raw bio-oil. This organic layer is the most
109 interesting to be upgraded for its potential use as biofuel due to its properties, both less
110 water and oxygen content, and higher heating value in comparison with the aqueous
111 layer. A detailed description of the experimental system used can be found elsewhere
112 [31-34].
113 Upgraded biofuels were produced via catalytic upgrading of raw bio-oil vapours (see
114 Supplementary Material) in a fixed-bed reactor at 450 ºC over several commercial and
115 lab-made zeolites, resulting in 70 biofuel samples. Details about the different zeolite
116 samples can be found in the Supplementary material (Tables S1-S2, Supplementary
117 material).
118
119 2.2. Biofuel characterization
120 The complete characterization of the organic phase for the upgraded biofuels was
121 carried out to determine elemental composition (C, H, O, S, N weight percent, as
122 received) (Carlo-Erba EA1108 according to UNE-EN ISO 16948:2015) [35], higher
123 heating value or HHV (IKA C-2000, according to UNE 164001 EX) [36] and water
124 content by Karl Fischer titration (Crison Titromatic KF, according to ASTM E203-16;
237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295
![Page 8: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/8.jpg)
6
125 weight percent) [17]. In addition, FTIR spectra of biofuels were obtained with a Bruker
126 Vertex 70 spectrometer corresponding to medium-infrared (MIR) covering the range
127 4000-400 cm-1. 50 μl liquid samples were applied on disposable real crystal IR sample
128 cards (KBr sample support substrate) with 15 mm of aperture provided by International
129 Crystal Laboratories. A Globar source and a deuterated L-alanine doped triglycine
130 sulphate detector were used. Each spectrum was acquired in transmission mode using
131 32 scans and a spectral resolution of 2 cm−1.
132
133 2.3. Chemometric methods and data analysis
134 2.3.1. Data pre-treatment for FTIR
135 Spectral data were imported from OPUS into the Unscrambler X 10.3 (Camo Inc., Oslo,
136 Norway) and they were converted to absorbance. All other spectral processing and
137 chemometric tools were performed using this program. The original data set contained
138 3735 variables (each one of the wavelengths obtained in the IR spectra), covering the
139 MIR region. Before building optimal PLS calibration models for each property, data
140 were centered by calculating the average value for each variable, and then, subtracting
141 this from each of the original variables. All variables were weighted at constant value
142 equal to 1 in order to avoid the influence of the different scales used for the variables.
143 Although different ranges of wavelengths were tested in order to assess the accuracy of
144 the PLS regression models obtained for each parameter under study, the best results for
145 the five bio-oil properties led to select the full MIR range: 4000-400 cm-1. Additionally,
146 combinations of three different spectral treatments were evaluated. The pre-processing
147 treatments eliminate or reduce the impact of the non-relevant spectral information and
148 often lead to simpler and more robust calibration models. These pre-processing
149 treatments were: a) Normalization to the highest peak, which attempts to correct for
296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354
![Page 9: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/9.jpg)
7
150 scaling differences, b) Standard Normal Variate (SNV), a row-oriented transformation
151 which centers and scales individual spectra, avoiding scatter effects from spectral data
152 and c) first derivative using Savitzky-Golay algorithms (SGolay) (polynomial order 2,
153 window with 11 points), which removes additive effects from the spectra, such as
154 baseline offsets or features. In this research, initial spectral treatments were limited to
155 the application of a SNV spectral treatment [37] (Fig. 1) since optimal results were
156 obtained, the same than in a previous work [38].
157
158 2.3.2. Partial least square regression models
159 Five PLS regression models were built, one for each biofuel property, where the 3735
160 variables of the IR spectra comprised between 4000 and 400 cm−1 for the biofuel oils
161 were used as predictor data (X) and the elemental analysis, water content and heating
162 value as the response variable (Y).
163 Samples were classified into two sets: half of samples (35 samples) were used to
164 develop the calibration models for each biofuel property, where full cross validation
165 (FCV), also known as leave-one-out cross validation, was selected to determine the
166 optimum number of latent variables (LV). The other half of the samples (35 samples)
167 was used as external prediction set to evaluate the predictive capacity of the model
168 between the predicted values from the calibration model and those from the reference
169 methods, using samples not used in the calibration model. Selection of the samples was
170 performed according to Kennard-Stone algorithm [39]. In general, external validation is
171 preferentially required [40] since full cross validation usually provides a too optimistic
172 assessment of the model predictive capacity.
173 The calibration models were performed by considering the correlation between the 35
174 IR spectral data and the five biofuel properties determined by the corresponding
355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413
![Page 10: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/10.jpg)
8
175 reference methods. The predictive capacity of the models obtained for each biofuel
176 property were performed by assessing the following criteria: the determination
177 coefficients (R2), which provide information regarding the fit of the model, the root
178 mean square errors (RMSE), for both calibration and prediction regression models [41]
179 and the bias, which can be interpreted as the average difference between the reference
180 value and the predicted value in the prediction set and refers to systematic errors.
181 Additionally, the ratio of performance to deviation (RPD) was calculated for each
182 model, which was determined by the coefficient between the standard deviation of the
183 input data and the standard error of the prediction (SD/SEP) [42, 43]. RMSE also
184 enabled to identify the presence of outliers that did not fit the model. A normal
185 probability plot of the Y-residuals of the model showing a fairly straight line with all
186 values within “+3SD” was also used to identify potential outliers. RMSE, SEP and bias
187 provide information about the accuracy, the precision and the trueness of the model,
188 respectively. The higher the RPD and the R2 values and the lower the RMSE, SEP and
189 bias values, the better the predictive capacity of the model.
190 A t-test for validation samples at 95% confidence level and degrees of freedom equal to
191 the number of prediction samples was carried out in order to establish whether there
192 were significant differences between the experimental and modelled results for each
193 biofuel property (Table S3, Supplementary material).
194 As estimation of the model uncertainty, a confidence interval (C.I.) was calculated by
195 the Eq. (1) when bias is small and standard deviation is unknown:
196 C.I. = (1)ȳ ± ≈ ȳ ± 197 where is the value predicted by the model and z-value must be replaced by a value ȳ198 from a Student´s t-distribution table which is estimated as 2 (1.960 is used for the 95%
414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472
![Page 11: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/11.jpg)
9
199 confidence interval) when N is high. This is an important measure, indicating the
200 uncertainty of the model.
201 The precision of each model, both absolute precision and relative precision, were
202 calculated according to the Eq (2) and (3) considering that bias values were close to
203 zero:
204 Absolute Precision = (2) . 205 Relative Precision (%) = *100 (3)
. ȳ206 According to Saeys et al. [44], RPD and R2 greater than 3.0 or 0.90, respectively, are
207 indicative of an excellent prediction. Values from 2.5 to 3.0 (RPD) and 0.82 to 0.90 (R2)
208 denote good prediction. RPD values between 2.0 and 2.5 and R2 values in the range
209 from 0.66 to 0.81 indicate approximate quantitative predictions. Finally, unsuccessful
210 predictions have RPD and R² values lower than 1.5 or 0.50, respectively, whilst poor
211 prediction have RPD and R² values of 1.5-2.0 and 0.50-0.65, respectively.
212
213 3. Results and discussion
214 3.1. Upgraded biofuel properties
215 A total of 70 upgraded biofuels (Table S4, Supplementary material) were produced. For
216 these samples, C, H and O elemental composition (wt.%), water content (wt.%) and
217 heating value (MJ kg-1) were determined according to reference methods mentioned in
218 Section 2.2. Biofuel characterization. Statistical descriptive of these properties for the
219 calibration and prediction sets are shown in Table 1.
220 Comparable values were observed for both groups of samples, confirming the reliability
221 of the selection algorithm. It is worth commenting that although C, H and O were the
222 main biofuel components, small amounts of nitrogen with a narrow compositional range
223 from 0.1 to 0.5 were also detected. According to the literature [45], standard deviation
473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531
![Page 12: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/12.jpg)
10
224 should be considered as a critical parameter in the development of any prediction model
225 since a small value together with a high standard error of the reference method hinder
226 the development of a good prediction model. Unfortunately, the presence of some
227 volatile components in the biofuels led to very relevant uncertainties during the
228 determination of the nitrogen content for these samples. This fact together with a low
229 variation for this parameter determined that nitrogen content would not be further
230 considered for modeling.
231 3.2. FTIR results
232 Basically, infrared energy can be absorbed at the atomic level originating alterations in
233 rotational-vibrational movements and giving rise to IR bands only if they cause a
234 change in the dipole moment of the molecule. These vibrational movements can be
235 classified as stretching (changes in the bond length) and bending (changes in bond
236 angles) and they can only reflect information of molecular functional groups but not
237 identify a specific chemical compound [46].
238 The FT-MIR spectra of the 70 biofuel samples are shown in Fig. 1 indicating a varied
239 organic composition. There were several broad peaks at different intensities related with
240 different functionalities in the biofuel and arising from activating molecular vibrations.
241 Two main zones were particularly remarkable in the spectra: 3600-2750 cm-1 and 1800-
242 700 cm-1. The first zone included the highest and broadest vibrations between 3600 and
243 3000 cm−1, indicating O-H stretching due to the presence of water and different
244 functional groups also related to phenols, alcohols [47], carboxylic acids, carbohydrates
245 and amino acids. Unsaturated and aromatic C-H bonds absorb from 3100 to 3000 cm−1
246 and aliphatic C-H bond absorbance bands were detected from 3000 to 2850 cm−1 [48]. A
247 broad and intense peak around 1700 cm−1 indicates C=O stretching vibration of free
248 carbonyl groups of aldehydes, ketones, carboxylic acids and esters and subsequent
532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590
![Page 13: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/13.jpg)
11
249 bands around 1650-1550 cm−1 represent C=C stretching vibrations caused by aliphatic
250 or aromatic structures. The spectral region of 1488-1400 cm−1 contains bands in the O-
251 H bending region, which were most probably associated with alcohols and carboxylic
252 acids [49]. The peak around 1260 cm−1 indicated C–O stretching peak (in acids, esters,
253 ethers and alcohols) and the peaks around 1030 cm−1 were associated with aromatic C-H
254 in plane bending. Aromatic rings could also be determined by the presence of C–H out
255 of plane bands [50] between 840 and 700 cm−1.
256 All the spectral information regarding different bonds seemed to be quite promising in
257 order to provide information regarding the biofuel properties: elemental composition
258 and water content. Because the heating value of biofuel oils is associated with the
259 content of C,H, O, N and S and their covalent carbon bonds, all these bands could also
260 be useful in determining this property.
261
262 3.3. Prediction of carbon content
263 A PLS calibration model was built from the set of samples selected by the Kennard-
264 Stone algorithm. It was observed that no more than two LV were required to explain
265 most of the variance (71.3% of the X variance and 86% of the Y variance)(Table 2).
266 The most important variables for the carbon content were shown in Fig. 2a) where the
267 weighted regression coefficients are shown for each LV as a function of the
268 wavenumbers. The highest coefficients involved in the modeling of this property were
269 included in the regions: 3080-2823; 2739-2723; 1809-1690; 1526-1506; 1263-1034 and
270 888-685 cm-1 that, as commented in section 3.2 FTIR results, were related with C-H
271 bonds, C=O and C-O stretching vibrations and C-H out of plane bands mainly caused
272 by aliphatic and aromatic structure.
591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649
![Page 14: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/14.jpg)
12
273 An assessment of the PLS regression prediction was carried out by means of the R2Pred
274 and the RMSEPred parameters (Table 2), which were 0.800 and 1.040 wt.%, respectively,
275 pointing out that the model could predict the carbon content of the upgraded biofuel
276 samples with approximate accuracy. RMSEPred and SEP were quite similar and the bias
277 value was close to zero, indicating that systematic errors were negligible. Accordingly,
278 RPD values higher than 2.0 and differences between R2Cal, R2
FCV and R2Pred lower than
279 0.2 [41] were obtained for the calibration and the prediction, showing that a robust
280 model was obtained with no overfitting.
281 However, it should be highlighted that the removal of one outlier in the prediction set
282 drove to a better model performance. Thus, a slight improvement in the R2Pred values
283 (Fig. 3) was obtained. This value increased from 0.800 to 0.836 whilst a decrease on the
284 RMSEPred=0.963 wt.% was observed.
285 Relative precision was also assessed for the prediction set based on one specific value
286 for the predicition set (RMSEPred), obtaining values lower than 3% for the carbon
287 content (Table 3). Despite the elemental analysis is an analytical technique providing
288 quite good results regarding precision, these results seem to be quite promising taking
289 into account that infrared with PLS is a rapid, non-destructive, reliable and low-cost
290 technique. However, it is remarkable to say that the use of only one value for RMSEPred
291 has some limitations and it can overestimate the model precision. In fact, when
292 repeatability of one specific sample was carried out by plotting five IR spectra, it was
293 observed that the relative precision based on the standard deviation was 0.11% (Table
294 3). Finally, a statistical analysis comparing the prediction results obtained by the PLS
295 regression model to those obtained by the standard analytical method was carried out
296 using a paired t-test at 95% level of confidence (Table S3, Supplementary material). T-
297 test results showed that differences between real and prediction values were not
650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708
![Page 15: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/15.jpg)
13
298 statistically significant and null hypothesis could not be rejected. Therefore, it could be
299 concluded that a PLS regression model could be accurately used to predict the carbon
300 content in upgraded biofuels obtained from the catalytic cracking of lignocellulosic
301 pyrolysis biofuels for a specific confidence interval between 73.74 and 77.51 wt.%
302 (Table 2).
303
304 3.4. Prediction of the hydrogen content
305 Regarding the hydrogen content, four LV were necessary to explain 88% of the X
306 variance and 83% of the Y variance in the calibration model. The weight regression
307 coefficients in the calibration model are shown in Fig. 2b) indicating that the most
308 interesting wavebands were the regions: 2985-2820 cm-1, 1524-1499 cm-1 and 1286-
309 1195 cm-1 associated with C-H, C=C and O-H stretching and bending vibrations. The
310 PLS regression model indicated that approximate predictions could be obtained for this
311 property, since R2Pred=0.757, RMSEPred=0.117 wt.% and RPD values higher than 2 were
312 obtained (Table 2). The removal of two outliers (Fig. 3) allowed increasing the R2Pred
313 value up to 0.815, again leading to better predictions and higher precision
314 (RMSEPred=0.101 wt.%). The relative uncertainties were also similar to those found for
315 carbon content prediction and lower than 3% (Table 3). Again, no statistically
316 significant differences were found between those values obtained by the standard
317 analytical method and the PLS regression model (Table S3, Supplementary material),
318 allowing the prediction of the hydrogen content with a reasonable degree of accuracy by
319 non-destructive techniques. Bias value was also close to zero, indicating similarity
320 between the RMSEPred and SEP and discarding systematic errors. In fact, very good
321 precision was obtained for the hydrogen content with values lower than 3% and 0.19%,
322 based on the RMSEPred and the standard deviation, respectively, for a specific sample.
709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767
![Page 16: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/16.jpg)
14
323 Herein, despite hydrogen content showed a narrow compositional range, this property
324 could be predicted with approximate accuracy within a confidence interval from 8.23 to
325 8.63 wt.% by our standard reference method, as shown in Table 2, allowing the
326 development of a consistent PLS regression model.
327
328 3.5. Prediction of the oxygen content
329 For this property, the calibration model was able to explain with two LV 71% and 87%
330 of the X and Y variances, respectively. The most important wavelengths associated with
331 this property were reflected in Fig. 2c) indicating that the regression coefficients were
332 affected by different spectral regions: 3600-2750 and 1800-500 cm-1. The region
333 between 3400-3680 cm-1, corresponding to O-H stretching vibrations, was mainly
334 caused by acid and/or alcohol structures. The bands between 3080-2821 cm-1 affected
335 negatively over the model and C-C stretching vibrations caused by aliphatic/aromatic
336 structure showed peaks at 1500-1527 cm-1 and C-O peaks at 1644-1800 cm-1 possibly
337 due to carboxylic acid/ester, aldehydes/ketone groups. At 1000-1300 cm-1, C-O groups
338 from acids, esters, ethers and alcohols and aromatic C-H bonds would be contributing to
339 the oxygen content modeling.
340 The oxygen content was the biofuel compositional element showing one of the best PLS
341 regression model performances regarding the fitting of the model. R2Pred= 0.812,
342 RMSEPred=1.087 wt.% and RPD=2.48 values were attained. According to these values,
343 approximate predictions could be obtained for this property. The removal of two
344 outliers (Fig. 3) allowed increasing the R2Pred up to 0.873, leading to good predictions.
345 Accordingly, a decrease of RMSEPred (0.910 wt.%) was obtained, probing that the
346 accuracy and the precision of the method was increased. Relative uncertainties were
347 lower than 15% (Table 3) for the prediction set, whilst an excellent precision was
768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826
![Page 17: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/17.jpg)
15
348 obtained based on the repeatability of IR spectra (0.48%). The confidence interval for
349 the prediction presented a wide compositional range from 14.00 to 17.57 wt.%. Results
350 of comparing the predicted data to those values obtained by the standard analytical
351 method did not provide any significant difference, pointing out that PLS regression
352 models could be used to foresee the oxygen content (Table S3, Supplementary material)
353 with notable accuracy.
354
355 3.6. Prediction of the water content
356 As shown in Table 2, two LV were required in the calibration model to explain 72% of
357 the X variance and 64% of the Y variance. The weight regression coefficients are
358 reflected in Fig. 2d) where the most important variables were associated basically with
359 different C-H, O-H, C=O, C-O functional groups. The highest coefficients were
360 obtained at 3519, 2870, 1516, 1271, 1034, 752 and 692 cm-1, respectively. The water
361 content was the property showing the lowest accuracy for the prediction model with
362 R2Pred and RMSEPred values of 0.705 and 0.577 wt.%, respectively (Table 2). These
363 parameters together with the narrow range of data modeled, pointed out that water
364 content could be only predicted with approximate quantitative accuracy. For this
365 property, relative uncertainties were the highest, about 26% (Table 3) within a
366 confidence interval for the prediction ranging from 3.36 to 5.62 wt.%. However, the
367 removal of two outliers in the prediction set significantly increased the R2Pred value up to
368 0.829 and good predictions were then obtained (Fig. 3). A RPD value higher than 3.0
369 was obtained and, again, no statistically significant differences were found between the
370 predictions with the PLS regression model and the experimental values, obtained by
371 Karl Fischer valorization (Table S3, Supplementary material).. Herein, for the first time
372 in the literature, it is shown that chemometric tools joint to medium-IR allow predicting
827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885
![Page 18: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/18.jpg)
16
373 the water content in upgraded lignocellulosic biofuels, as a fast screening tool instead of
374 the standard Karl Fischer valorization method.
375
376 3.7. Prediction of the heating value
377 The calibration model established two LV to explain 71% of the X variance and 90% of
378 the Y variance. The most important variables describing the heating value were shown
379 in Fig. 2e) where the weight regression coefficients were plot over wavenumbers. It is
380 remarkable the similarity of the important wavelengths (3505, 2922, 2855, 1712, 1516,
381 1271, 1035, 752 and 692 cm-1) with the ones obtained for the carbon content (Fig. 2a))
382 indicating that the heating value is quite depending on the carbon content for these
383 biofuel samples, as expected.
384 Heating value was the fuel property showing the best fitting parameters, since R2Pred and
385 RMSEPred values were 0.836 and 0.599 MJ kg-1, respectively. A good prediction was
386 obtained, which was notably improved after the removal of one outlier (Fig. 3),
387 increasing the R2Pred = 0.874 and decreasing the RMSEPred=0.539 MJ kg-1. Relative
388 uncertainties were also lower than 4% with values of 0.10% for the repeatability of IR
389 spectra (Table 3). RPD value was 2.56 and as for the other modeled properties, no
390 statistically significant differences were found between the prediction by the external
391 validation and the real values obtained by the standard analytical technique
392 (calorimetric bomb). Summarizing, this PLS regression model could be successfully
393 used to predict the heating value with notable accuracy, especially taking into account
394 that this is the first time that the heating value of upgraded biofuels has been predicted
395 based on chemometric tools, using PLS regression and MIR spectra. These results are
396 quite promising to advance in the use of non-destructive and sensing techniques that
397 involve more rapid analysis and less inversion costs than other traditional procedures in
886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944
![Page 19: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/19.jpg)
17
398 the characterization of second generation biofuels obtained from the catalytic upgrading
399 of pyrolysis lignocellulosic bio-oils.
400
401 4. Conclusions
402 The search of techniques that allow determining specific properties of second
403 generation biofuels produced from lignocellulosic biomass in a fast way are of great
404 concern. In this research, the coupling of mid-infrared with chemometric tools based on
405 partial least square regressions have proven to be a promising tool in order to predict
406 quantitatively some relevant properties (C, O, H, water content and heating value)
407 associated to upgraded biofuels obtained from the catalytic cracking of pyrolysis
408 lignocellulosic bio-oils. Robust models were built for each property, always achieving
409 good predictions, R2Pred ≥0.82. In this way, an easy assessment of these properties could
410 be performed by determining the MIR spectra of upgraded biofuels, avoiding more
411 tedious traditional procedures that require more time and expenses. These results could
412 be considered in a first approach as a promising and useful tool in order to perform
413 quality controls of second generation biofuels production technologies. More research
414 should be conducted to increase the number of upgraded bio-oils and to study how the
415 application of other catalysts could affect the prediction of the PLS models for these
416 five properties.
417418 Appendix A. Supplementary material419420 Supplementary material related to this article can be found at:
421 Acknowledgements
422 Authors would like to thank the Spanish MINECO and European Union FEDER funds
423 for supporting this work (projects CTQ2012-37984-C02-01 and ENE2015-68320-R).
424
9459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003
![Page 20: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/20.jpg)
18
425 References
426 [1] A. Friedl, E. Padouvas, H. Rotter, K. Varmuza, Prediction of heating values of
427 biomass fuel from elemental composition, Analytica Chimica Acta 544 (2005) 191-198.
428 [2] W. Laursen, Students take a green initiative, Chem. Eng. (2006) 32-34.
429 [3] S.N. Naik,V.V. Goud, P.K. Rout, A.K. Dalai, Production of first and second
430 generation biofuels: A comprehensive review. Renew. Sustainable Energy Rev. 14(2)
431 (2010) 578-597.
432 [4] A.V. Bridgwater, D. Meier, D. Radlein, An overview of fast pyrolysis of biomass,
433 Org. Geochem. 30 (1999)1479.
434 [5] D. Meier, O. Faix, State of the art of applied fast pyrolysis of lignocellulosic
435 materials––a review, Biosource Technol. 68(1) (1999) 71–77.
436 [6] D. Mohan, Jr. CU. Pittman, P.H. Steele, Pyrolysis of wood/biomass for bio-oil: A
437 critical review, Energy Fuels 20 (2006) 848-889.
438 [7] L. Qiang, L. Wen-Zhi, Z. Xi-Feng, Overview of fuel properties of biomass fast
439 pyrolysis oils, Energy Convs. Manage. 50 (2009) 1376-1383.
440 [8] I. Graça, J.M. Lopes, H.S. Cerqueira, M.F. Ribeiro, Bio-oils upgrading for second
441 generation biofuels, Ind. Eng. Chem. Res. 52 (2013) 275-287.
442 [9] S. Kathiravale, M.N. Mudhd-Yunus, K. Sopian, A.H. Samsuddin, R.A. Rahman,
443 Modeling the heating value of municipal solid waste, Fuel 82(9) (2003) 1119-1125.
444 [10] CEN/TS 14918: 2005. Solid biofuels. Method for the determination of calorific
445 value.
446 [11] R.S. Jessup, Precise Measurement of Heat Combustion with a Bomb Calorimeter,
447 U.S. Bureau of Standards Monograph No. 7, US Government Printing Office,
448 Washington, D.C., 1970.
10041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062
![Page 21: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/21.jpg)
19
449 [12] C. Sheng, J.L.T. Azevedo, Estimating the higher heating value of biomass fuels
450 from basic analysis data, Biomass Bioenergy 28 (2005) 499–507.
451 [13] R. Saidur, E.A. Abdelaziz, A. Demirbas, M.S. Hossain, S. Mekhilef, A review on
452 biomass as a fuel for boilers, Renew. Sustain. Energy Rev. 15 (2011) 2262–89.
453 [14] J.F. Saldarriaga, R. Aguado, A. Pablos, M. Amutio, M. Olazar, J. Bilbao, Fast
454 characterization of biomass fuels by thermogravimetric analysis (TGA), Fuel 140
455 (2015) 744–51.
456 [15] CEN/TS 15104: 2005. Solid biofuels. Determination of total content of carbon,
457 hydrogen and nitrogen. Instrumental methods.
458 [16] K. Fischer, Karl (1935). Neues verfahren zur maßanalytischen bestimmung des
459 wassergehaltes von flüssigkeiten und festen körpern. Angew. Chem. 48 (26): 394–396,
460 doi:10.1002/ange.19350482605.
461 [17] ASTM E203 – 16. Standard test method for water using volumetric Karl Fischer
462 titration.
463 [18] D.B. Dahlberg, S.M. Lee, S.J. Wenger, J.A. Vargo, Classification of vegetable oils
464 by FT-IR. Appl. Spectrosc. 51 (1997) 1118–1124.
465 [19] Z. Droussi, V. D’Orazio, M.R. Provenzano, M. Hafidi et al. Ouatmane Study of the
466 biodegradation and transformation of olive-mill residues during composting using FTIR
467 spectroscopy and differential scanning calorimetry, J.Hazard. Mater. 164 (2009) 1281–
468 1285.
469 [20] W. Terouzi, M. De Luca, A. Bolli, A. Oussama, M. Patumi, G. Ioele, G. Ragno, A
470 discriminant method for classification of Moroccan olive varieties by using direct FT-IR
471 analysis of the mesocarp section,Vib. Spectro. 56 (2011) 123–128.
10631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121
![Page 22: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/22.jpg)
20
472 [21] W. Terouzi, Z. Ait Yacine, F. Kzaiber, A. Boulli, A. Oussama, Kernel olive:
473 Marker of tracing the variety by using the FTIR spectra and chemometrics. Int. J.
474 Chemtech. Appl. 3(4) (2014) 1-13.
475 [22] C.E. Miller, Chemometrics for on-line spectroscopy applications theory and
476 practice, J. Chemom. 14(5–6) (2000) 513–528.
477 [23] K.H. Esbensen, Multivariate data analysis-in practice. An introduction to
478 multivariate data analysis and experimental desing. Oslo: CAMO; 2000.
479 [24] B.G.M. Vandeginste, D.L. Massart, L.M.C. Buydens, S.D.E. Jong, P.J. Lewi, J.
480 Smeyers-Verbeke, Handbook of Chemometrics and Qualimetrics: Part B, Elsevier,
481 Amsterdam, 1998, p 88.
482 [25] P.R. Wiley, G.J. Tanner, P.M. Chandler, R.S. Anderssen, Molecular classification
483 of barbey (Hordeum vulgare L.) mutants using derivative NIR spectroscopy, J. Agric.
484 Food Chem. 57 (2009) 4042-4050.
485 [26] S. Vonhoff, J. Condliffe, H. Schiffter, Implementation of an FTIR calibration curve
486 for fast and objective determination of changes in protein secondary structure during
487 formulation development, J. Pharm. Biomed. Anal. 51 (2010) 39-45.
488 [27] D.T. Chadwick, K.P. McDonnell, L.P. Brennan, C.C. Fagan, C.D. Everard,
489 Evaluation of infrared techniques for the assessment of biomass and biofuel quality
490 parameters and conversion technology processes: A review, Renew. Sustainable Energy
491 Rev. 30 (2014) 672-681.
492 [28] G.D. Gillespie, C.D. Everard, K.P. McDonnell, Prediction of biomass pellet quality
493 indices using near infrared spectroscopy, Energy 80 (2015) 582-588.
494 [29] Al-Ghouti M, Al-Degs Y, Mustafa F. Determination of hydrogen content, gross
495 heat of combustion and net heat of combustion of diesel fuel using FTIR spectroscopy
496 and multivariate calibration. Fuel 2010; 89: 193-01.
11221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180
![Page 23: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/23.jpg)
21
497 [30] Y. Felkel, N. Dörr, F. Glatz, K. Varmuza, Determination of the total acid number
498 (TAN) of used gas engine oils by IR and chemometrics applying a combined strategy
499 for variable selection, Chemometr. Intell. Lab. 101 (2010) 14–22.
500 [31] A. Veses, M. Aznar, J.M. López, M.S. Callén, R. Murillo, T. García, Production of
501 upgraded bio-oils by biomass catalytic pyrolysis in an auger reactor using low cost
502 materials, Fuel 141 (2015) 17-22.
503 [32] A. Veses, B. Puértolas, M.S. Callén, T. García, Catalytic upgrading of biomass
504 derived pyrolysis vapors over metal-loaded ZSM-5 zeolites: Effect of different metal
505 cations on the bio-oil final properties, Microp. Mesop. Mat. 209 (2015) 189-196.
506 [33] A. Veses, B. Puértolas, J.M. López, M.S. Callén, B. Solsona, T. García, Promoting
507 Deoxygenation of Bio-Oil by Metal-Loaded Hierarchical ZSM-5 Zeolites, ACS
508 Sustainable Chem. Eng. 4(3) (2016) 1653–1660.
509 [34] A. Veses, M. Aznar, I. Martínez, J.D. Martínez, J.M. López, M.V. Navarro, M.S.
510 Callén, R. Murillo, T. García, Catalytic pyrolysis of Wood biomass in an auger reactor
511 using calcium-based catalysts, Bioresour. Technol. 162 (2014) 250-258.
512 [35] UNE-EN ISO 16948:2015. Solid biofuels. Determination of total content of
513 carbon, hydrogen and nitrogen.
514 [36] UNE 164001 EX. Solid Biofuels. Method for the determination of calorific value.
515 [37] V.R. Nadadoor, H. De la Hoz Siegler, S.L. Shah, W.C. McCaffrey, A. Ben-Zvi,
516 Online sensor for monitoring a microalgal bioreactor system using supportvector
517 regression, Chemometr. Intell. Lab. 110 (2012) 38-48.
518 [38] T. García, A. Veses, J.M. López, B. Puértolas, J. Pérez-Ramírez, M.S. Callén,
519 Determining bio-oil composition via chemometric tools based on infrared spectroscopy,
520 ACS Sustain. Chem. Eng. 5(10) (2017) 8710-8719.
11811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239
![Page 24: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/24.jpg)
22
521 [39] R.W. Kennard, L.A. Stone, Computer aided design of experiments, Technometrics
522 11(1) (1969) 137-148.
523 [40] G.C.S. Smith, S.R. Seaman, A.M. Wood, P. Royston, I.R. White, Correcting for
524 optimistic prediction in small data sets, Am. J. Epidemiol. 180(3) (2014) 318-324.
525 [41] D.L. Sills, J.M. Gossett, Using FTIR spectroscopy to model alkaline pretreatment
526 and enzymatic saccharification of six lignocellulosic biomasses, Biotechnol. Bioeng.
527 109(4) (2012) 894-903.
528 [42] P.C. Williams, D.C. Sobering, How do we do it: a brief summary of the methods
529 we use in developing near infrared calibrations, in: A.M.C. Davies , P. Williams (Eds.),
530 Near Infrared Spectroscopy: the future waves, NIR Publications, Chichester, UK, 1996,
531 pp. 185-188.
532 [43] T. Fearn, Assessing calibrations: SEP, RPD, RER, and R2, NIR News13 (2002)
533 12–14.
534 [44] W. Saeys, A.M. Mouazen, H. Ramon, Potential for onsite and online analysis of
535 pig manure using visible and near infrared spectroscopy, Biosyst. Eng. 91(2005) 393–
536 402.
537 [45] J. Mata Sánchez, J.A. Pérez Jiménez, M.J. Díaz Villanueva, A. Serrano, N. Núñez,
538 J. López Giménez, Assessment of near infrared spectroscopy for energetic
539 characterization of olive byproducts, Renew. Energy 74 (2015) 599-605.
540 [46] C.C. Sousa, J.A. Lopes, Infrared spectroscopy detection coupled to chemometrics
541 to characterize foodborne pathogens at a subspecies level, in: D. Granato (Ed.),
542 Mathematical and Statistical Methods in Food Science and Technology, John Wiley &
543 Sons, Ltd, Chichester, UK, 2014, DOI: 10.1002/9781118434635.ch20.
12401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298
![Page 25: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/25.jpg)
23
544 [47] M.I. Schnitzer, C.M. Monreal, G.A. Facey, P.B. Fransham, The conversion of
545 chicken manure to biooil by fast pyrolysis I, Analysis of biooils by FTIR and NMR
546 spectroscopy, J. Environ. Sci. Health, Part B 42 (2007) 71–77.
547 [48] C. Coury, A.M. Dillner, A method to quantify organic functional groups and
548 inorganic compounds in ambient aerosols using attenuated total reflectance FTIR
549 spectroscopy and multivariate chemometric techniques, Atmos. Environ. 42 (2008)
550 5923–5932.
551 [49] M. Asadieraghi, W.M.A. Wan Dau, In-depth investigation on thermochemical
552 characteristics of palm oil biomasses as potential biofuel sources, J. Anal. Appl. Pyrol.
553 115 (2015) 379–391.
554 [50] P.K. Kanaujia, Y.K. Sharma, M.O. Garg, D. Tripathi, R. Singh, Review of
555 analytical strategies in the production and upgrading of bio-oils derived from
556 lignocellulosic biomass, J. Anal. Appl. Pyrol. 105 (2014) 55–74.
557
12991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357
![Page 26: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/26.jpg)
24
Table 1. Statistical descriptive of the different properties for the set of samples (N= number of samples; Cal.= Calibration; Pred.= Prediction, KS= Kennard-Stone algorithm).
Samples N C(wt.%)
O (wt.%)
H(wt.%)
N(wt.%)
WaterContent (wt.%)
Heatingvalue (MJ kg-1)
Cal. KS 35Mean 75.1 16.4 8.38 0.2 4.72 34.5SDCal. 2.5 2.7 0.25 0.1 1.26 1.6
Maxim 78.2 21.0 8.72 0.5 7.04 36.4Minim 70.8 13.2 7.89 0.1 3.11 31.6
Pred. KS
35
Mean 75.4 16.0 8.40 0.2 4.45 34.8SDPred 2.3 2.4 0.22 0.1 1.04 1.4Maxim 78.8 20.9 8.75 0.4 7.24 36.9Minim 70.9 12.3 8.01 0.1 2.79 31.9
558 N= number of samples; SDCal.= Standard deviation of the calibration set; SDPred.= Standard deviation 559 of the prediction set; wt.%=weight percent; Maxim=maximum value and Minim=minimum value.560
561
562
563
13581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416
![Page 27: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/27.jpg)
25
Tabl
e 2.
Sum
mar
y of
the
parti
al le
ast s
quar
es re
gres
sion
s m
odel
s de
velo
ped
to p
redi
ct u
ltim
ate
anal
ysis
(wt.%
), w
ater
con
tent
(wt.%
) an
d he
atin
g va
lue
(MJ
kg-1
) (N
= nu
mbe
r of s
ampl
es, L
V=l
aten
t var
iabl
es, R
MSE
=roo
t mea
n sq
uare
err
or, S
EC/S
EP=s
tand
ard
erro
r of
cal
ibra
tion
or p
redi
ctio
n, R
PD=r
atio
of
perf
orm
ance
to
devi
atio
n, R
2 = d
eter
min
atio
n co
effic
ient
, C
.I.=c
onfid
ence
in
terv
al).
Dep
ende
ntW
avel
engt
hN
Mod
el
type
LVV
aria
nce
Bia
sR
MSE
SEC
/SEP
RPD
R2
C.I.
C40
00-4
0035
Cal
.2
71.3
86.0
-1.1
x10
-20.
934
0.94
72.
600.
860
73.1
9-76
.85
35Pr
ed.
2-
-2.
2 x1
0-11.
040
1.12
32.
190.
800
73.6
0-77
.68
34Pr
ed.
23.
0 x1
0-10.
963
1.11
02.
220.
836
73.7
4-77
.51
H40
00-4
0035
Cal
.4
87.7
82.5
1.8
x10-3
0.10
40.
106
2.32
0.82
58.
18-8
.59
35Pr
ed.
4-
-1.
3 x1
0-20.
117
0.12
02.
050.
757
8.19
-8.6
533
Pred
.4
3.3
x10-2
0.10
10.
119
2.06
0.81
58.
23-8
.63
O40
00-4
0035
Cal
.2
71.2
87.0
1.1
x10-2
0.97
10.
986
2.70
0.87
014
.47-
18.2
835
Pred
.2
--
-2.6
x10
-11.
087
1.07
12.
480.
812
13.5
8-17
.83
33Pr
ed.
22.
8 x1
0-10.
910
0.87
73.
400.
873
14.0
0-17
.57
Wat
er40
00-4
0035
Cal
.2
72.0
63.7
-9.5
x10
-30.
770
0.78
21.
610.
637
3.21
-6.2
335
Pred
.2
--
4.4
x10-2
0.57
70.
590
2.14
0.70
53.
36-5
.62
33Pr
ed.
25.
6 x1
0-20.
416
0.41
83.
020.
829
3.60
-5.2
3H
eatin
g va
lue
4000
-400
35C
al.
271
.290
.2-3
.3 x
10-3
0.50
20.
509
3.11
0.90
233
.49-
35.4
635
Pred
.2
--
1.1
x10-1
0.59
90.
639
2.47
0.83
633
.69-
36.0
434
Pred
.2
1.6
x10-1
0.53
90.
618
2.56
0.87
433
.80-
35.9
1
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
![Page 28: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/28.jpg)
26
Table 3 Relative precision (%) at 95 % confidence level in the calibration and prediction sets obtained with the PLS regression models.
Properties Relative precision
Calibration N=35C 2.46H 2.41O 11.67Water content 27.82Heating value 2.93
Prediction N=35
C 2.69 (0.11)H 2.78 (0.19)O 13.85 (0.48)Water content 25.70 (1.90)Heating value 3.44 (0.10)In brackets the relative precision at 95% confidence level based on the repeatability of five IR
spectra of the same sample calculated according to standard deviation.
14581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516
![Page 29: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/29.jpg)
27
4000
3500
3000
2500
2000
1500
1000
500
0.0
0.5
1.0
1.5
2.0
Absorbance (a.u.)
Wav
enum
bers
(cm
-1)
Cal
ibra
tion
sam
ples
4000
3500
3000
2500
2000
1500
1000
500
-1.5
-1.0
-0.50.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
Absorbance (a.u.)
Wav
enum
bers
(cm
-1)
Cal
ibra
tion
sam
ples
SN
V
4000
3500
3000
2500
2000
1500
1000
500
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
Absorbance (a.u.)
Wav
enum
bers
(cm
-1)
Pred
ictio
n sa
mpl
es
4000
3500
3000
2500
2000
1500
1000
500
-1.5
-1.0
-0.50.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
Absorbance (a.u.)
Wav
enum
bers
(cm
-1)
Pred
ictio
n sa
mpl
es S
NV
Fig.
1FT
IR s
pect
rum
of
the
calib
ratio
n an
d pr
edic
tion
biof
uel
sam
ples
cor
resp
ondi
ng t
o th
e M
IR r
egio
n: 4
000-
400
cm-1
w
ithou
t any
spec
tral t
reat
men
t (le
ft si
de) a
nd a
fter s
tand
ard
norm
al v
aria
te sp
ectra
l tre
atm
ent (
SNV
)(rig
ht si
de).
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
![Page 30: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/30.jpg)
28
a)
4000 3500 3000 2500 2000 1500 1000 500-0.06
-0.05
-0.04
-0.03
-0.02
-0.01
0.00
0.01
0.02
0.03R
egre
ssio
n co
effic
ient
s (BW
)
Wavenumbers (cm-1)
Factor 1Factor 2Carbon content
b)
4000 3500 3000 2500 2000 1500 1000 500-0.006
-0.005
-0.004
-0.003
-0.002
-0.001
0.000
0.001
0.002
0.003
0.004
0.005
0.006
0.007
Reg
ress
ion
coef
ficie
nts (
BW)
Wavenumbers (cm-1)
Factor 1Factor2Factor 3Factor 4
Hydrogen content
c)
4000 3500 3000 2500 2000 1500 1000 500
-0.03
-0.02
-0.01
0.00
0.01
0.02
0.03
0.04
0.05
0.06
Reg
ress
ion
coef
ficie
nts (
BW)
Wavenumbers (cm-1)
Factor 1Factor 2Oxygen content
15581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616
![Page 31: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/31.jpg)
29
d)
4000 3500 3000 2500 2000 1500 1000 500-0.02
-0.01
0.00
0.01
0.02
0.03
Reg
ress
ion
coef
ficie
nts (
BW)
Wavenumbers (cm-1)
Factor 1Factor 2Water content
e)
4000 3500 3000 2500 2000 1500 1000 500-0.04
-0.03
-0.02
-0.01
0.00
0.01
0.02
Reg
ress
ion
coef
ficie
nts (
BW)
Wavenumbers (cm-1)
Factor 1Factor 2Heating value
Fig. 2. Weighted regression coefficients of the FTIR-PLS calibration model of: a) C, b) H, c) O, d) water content and e) heating value in upgraded biofuel samples.
16171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675
![Page 32: JAAP 2017 826 Revision 1 V0 - Consejo Superior de](https://reader034.vdocument.in/reader034/viewer/2022042802/62683354f5d6612386304ba3/html5/thumbnails/32.jpg)
30
7071727374757677787980
70 72 74 76 78 80
Pred
icte
d C
(wt.%
)
Reference C (wt.%)
R2=0.800RMSEPred=1.040SEP=1.123Bias=0.219
71
72
73
74
75
76
77
78
79
80
70 72 74 76 78 80
Pred
icte
d C
(wt.%
)
Reference C (wt.%)
R2=0.836RMSEPred=0.963SEP=1.110Bias=0.300
7.87.9
88.18.28.38.48.58.68.78.8
8 8.2 8.4 8.6 8.8 9
Pred
icte
d H
(wt.%
)
Reference H (wt.%)
R2=0.757RMSEPred=0.117SEP=0.120Bias=0.013
7.9
8
8.1
8.2
8.3
8.4
8.5
8.6
8.7
8.8
8 8.2 8.4 8.6 8.8 9
Pred
icte
d H
(wt.%
)Reference H (wt.%)
R2=0.815RMSEPred=0.101SEP=0.119Bias=0.033
0
5
10
15
20
25
10 12 14 16 18 20 22
Pred
icte
d O
(wt.%
)
Reference O (wt.%)
R2=0.812RMSEPred=1.087SEP=1.071Bias=-0.256
0
5
10
15
20
25
10 12 14 16 18 20 22
Pred
icte
d O
(wt.%
)
Reference O (wt.%)
R2=0.873RMSEPred=0.910SEP=0.877Bias=-0.285
2
3
4
5
6
7
8
2 4 6 8
Pred
icte
d w
ater
(wt.%
)
Reference water (wt.%)
R2=0.705RMSEPred=0.577SEP=0.590Bias=0.044
2
3
4
5
6
7
8
2 4 6 8
Pred
icte
d w
ater
(wt.%
)
Reference water (wt.%)
R2=0.829RMSEPred=0.416SEP=0.418Bias=-0.056
31
32
33
34
35
36
37
38
32 33 34 35 36 37
Pred
icte
d h
eat.
valu
e (M
J/kg
)
Reference heat. value (MJ/kg)
R2=0.836RMSEPred=0.599SEP=0.639Bias=0.112
3031323334353637383940
32 33 34 35 36 37
Pred
icte
d h
eat.
valu
e (M
J/kg
)
Reference heat. value (MJ/kg)
R2=0.874RMSEPred=0.539SEP=0.618Bias=0.163
Fig. 3 Plot of the prediction set (left graph for all samples, N=35 samples, right graph after removing outliers) for the different properties obtained by PLS regression model.
16761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734