jaap 2017 826 revision 1 v0 - consejo superior de

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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.%)


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.%)


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.%)


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.%)


2

3

4

5

6

7

8

2 4 6 8

Pred

icte

d w

ater

(wt.%

)

Reference water (wt.%)


2

3

4

5

6

7

8

2 4 6 8

Pred

icte

d w

ater

(wt.%

)

Reference water (wt.%)


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)


3031323334353637383940

32 33 34 35 36 37

Pred

icte

d h

eat.

valu

e (M

J/kg

)

Reference heat. value (MJ/kg)


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

jaap 2017 826 revision 1 v0 - consejo superior de

Documents