- Email: [email protected]
Prediction of gross calorific value and ash content of woodchip samples by means of FT-NIR spectroscopy
Fuel Processing Technology 169 (2018) 77–83
Contents lists available at ScienceDirect
Fuel Processing Technology journal homepage: www.elsevier.com/locate/fuproc
Prediction of gross calorific value and ash content of woodchip samples by means of FT-NIR spectroscopy
MARK
M. Mancinia, Å. Rinnanb, A. Pizzia, G. Toscanoa,⁎ a b
Agricultural, Food and Environmental Sciences, Università Politecnica delle Marche, via Brecce Bianche, 60131 Ancona, Italy Department of Food Science, Faculty of Science, University of Copenhagen, Rolighedsvej 26, 1958 Frederiksberg C, Denmark
A R T I C L E I N F O
A B S T R A C T
Keywords: PLS Variables selection rPLS Biomass quality Woodchip
The use of woodchip, and biofuels in general, is a fundamental step towards the transition from fossil fuel to renewable energy. The growth in the demand for wood fuels and the inherent variability in the properties of woody material lead to the need to verify its quality. EN ISO 17225-4 divides woodchip in different quality classes according to chemical-physical parameters and quality attributes. In this study, we have coupled near infrared spectroscopy with Partial Least Square regression to model gross calorific value and ash content of woodchip samples. Moreover, variables selection methods were tested in order to improve the models and get better prediction. Gross calorific value and ash content were predicted with a standard error of 234 J/g and 0.44%, respectively. The results indicate that the models could be used in screening applications and near infrared spectroscopy is a promising tool in the evaluation of biomass quality.
1. Introduction In recent years, the interest in bioenergy as an alternative to fossil energy is increasing because of the continuously growing energy demand, the decreasing availability of fossil fuels and the need for a reduction of environmental impact. In order to reach these targets European policies is aiming to promote the use of renewable energy sources. Woody biomass is such a source of energy, it is present more or less everywhere, is available in many forms and can be easily stored, especially in comparison with other energy sources [1]. In particular, woodchip is really appreciated because it consists of homogeneous particles with a specific size and it guarantees benefits in terms of increased load density and handling quality [2,3]. In different European countries, and in Italy as well, the number of power plants fueled with woodchip is increasing and accordingly also the demand for wood fuels. As a consequence, the biomass quality could experience a decrease and need to be analysed [4]. Moreover, it is known that there is an inherent variability in the properties of woody material that is influenced by many factors [5,6] and this leads to the need to employ quality standards in order to check the quality of the product. CEN/TC 335 has established a number of standards to ensure biomass quality. Chemical-physical parameters and quality features (e.g. origin and source) divide the biomass in different quality classes. The
⁎
identification and characterization of chemical composition of a solid fuel is the most important step during the investigation and application of such fuel. This composition is a unique code that characterizes and determines the properties, quality, application perspectives and environmental problems related to any fuel [7]. EN ISO 17225–4 [8] is the quality standard for graded woodchip for residential, small commercial and public building applications and provides limits for three different woodchip quality classes: A1, A2 and B. In particular, calorific value is an essential parameter in the specification of biomass quality to set the price of the product. Moreover it is an important characteristic for the planning and control of power plants using biomass fuel [9]. Ash content (Ac) influences combustion efficiency and may cause cleaning and combustion problems, such as slagging in furnaces, fouling of heat exchanger surfaces, corrosion in the combustion device [4,10,11]. Ac is also the most important discriminant parameter among the woodchip quality classes that are related to a maximum Ac of 1.0, 1.5 and 3.0% respectively. Gross calorific value (GCV) and Ac are normally determined by traditional laboratory analysis, but the process is tedious, expensive and requires specialized experts. As a consequence, it is necessary to develop a technique that is rapid, economic and simple. A good candidate could be Near Infrared (NIR) spectroscopy which is already widely used for quantitative and qualitative purposes in different sectors, i.e. pharmacy, food and agricultural industries.
Corresponding author. E-mail address: [email protected] (G. Toscano).
http://dx.doi.org/10.1016/j.fuproc.2017.09.021 Received 23 May 2017; Received in revised form 19 September 2017; Accepted 20 September 2017 0378-3820/ © 2017 Elsevier B.V. All rights reserved.
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al.
701 Leco). The Acad was determined as the ratio between the residue remaining after the sample was heated in air at 550 ± 10 °C to initial biomass. The average Acad was calculated based on two measurements per sample. The ash content on a dry basis (Acdb) was obtained for each sample by multiplying Acad by its dry matter content in percentage. A subset of samples (n = 86) was selected for GCV analysis. Gross calorific value of air dried ground material (GCVad) was determined in a dynamic mode at 25 °C using a bomb calorimeter (mod.C2000 basic, IKA). The calorimeter was calibrated using a benzoic acid standard (IKA Benzoic Acid C723). The analyses were performed in duplicate for each sample. As for Ac, gross calorific value on a dry basis (GCVdb) was obtained for each sample by multiplying GCVad by the dry matter content in percentage.
A literature research, shows that only a limited number of studies has been performed on the use of NIR spectroscopy in the field of biomass control. Fagan et al. [12] and Gillespie et al. [13] examined the possibility to employ NIR spectroscopy and chemometrics to predict moisture, ash, carbon contents and gross calorific value of Miscanthus, willow and herbaceous energy grasses. The possibility to predict calorific value, moisture and ash content was studied also by Lestander and Rhen [14] on Norway spruce (Picea abies (L.) Karst.). Moreover, studies have shown the potential of NIR spectroscopy to predict calorific value in dedicated bioenergy crops [15], in Leucaena leucocephala pellets [9] and in straw [16]. Saha et al. [17] investigated the possibility to predict calorific values, moisture, ash, carbon, nitrogen, and sulfur content of pine tree biomass using near infrared spectroscopy. Lower heating value and elemental composition were also examined by Pasom et al. using partial least squares regression and considering bamboo samples [18]. The ash content was also investigated in wheat straw by Bruun et al. with good results [19]. All the prediction models developed in the aforementioned studies were based on a specific kind of biomass or wood species. To the authors' knowledge, there are no works based on woodchip samples taken directly from the market, without having information about wood typology. This approach allows the development of more robust and reliable prediction models, which could be employed on a national scale. Furthermore, in the solid biofuels sector, no variable selection methods were applied to NIR models for the improvements of the models performance. The objective of this study is to develop models based on Fourier Transform Infrared (FT-NIR) Spectroscopy and multivariate analysis (Partial Least Squares) for the prediction of GCV and Ac of woodchip samples. In order to achieve this, a large number of samples from several power plants, representative of the national scene, were prepared and analysed by FT-NIR spectroscopy. The multivariate regression models were further elaborated using chemometric techniques, such as variables selection combined with data pre-processing, in order to get better prediction models.
2.3. Near-infrared spectroscopy The near infrared spectra were collected using a FT-NIR spectrophotometer (mod. Antaris II, Thermo Fisher Scientific Inc., USA) equipped with a halogen lamp as a source and an InGaAs detector. The samples were acquired in diffuse reflectance mode using an integrating sphere and were kept in rotation during the acquisition by means of a sample cup spinner to increase the representativeness of the material. Each spectrum has been computed as an average of 32 successive scans acquired at a wavelength range from 10,000 to 4000 cm− 1 and with the spectral resolution of 8 cm− 1. All measurements have been performed at room temperature (18–20 °C) and each sample is recorded in duplicate. A blank spectrum was acquired every hour to exclude the signals not associated to the sample, but to the instrument or environment. 2.4. Data processing and Partial Least Square Regression NIR spectroscopy requires chemometrics to extract as much relevant information as possible from the analytical data [20]. Partial Least Square Regression (PLS) is a common method applied in spectroscopy for quantitative analysis. It finds the relationship between a y-value the parameter to be quantified - and the spectral data matrix, maximizing the covariation between them. PLS finds a new smaller set of variables, called latent variables (LVs), that are linear combinations of the spectral data and relevant for the determination of the parameters of interest [21]. In this study, PLS models were calculated to predict Ac and GCV of woodchip samples both air dried and on the dry basis. Prior to PLS regression, in order to minimize the effect of baseline shifts and noise, the spectra were pre-processed [22]. Different pre-treatments including Standard Normal Variate (SNV), first and second derivative spectra (Savitzky-Golay [23] with 13 or 21 smoothing points and 2nd order polynomial) were applied. As no replicate outliers were found, each sample was averaged across the replicates. Residual vs. leverage, observed vs. predicted response and PLS score plots were utilized in order to identify possible sample outliers. PLS regression models were validated using Venetian blind-cross validation (5 segments). In this type of validation, n validation models are created from the original data set in order to assess the performance of the prediction model. For each validation model, the test set is obtained taking out samples at every n position in the original data set, while the remaining samples represent the training set which is used to build up the model. Therefore, a group of samples is left out from the calibration data set and the model is calibrated on the remaining data point. Then the values for the test set are predicted and the prediction residuals are calculated. The procedure is repeated until each subset has been left out once. As a result, all prediction residuals are combined to calculate validation residual variance and the root mean square error of
2. Materials and methods 2.1. Sample collection and preparation In this study a total of 125 woodchip samples were collected and analysed by the Biomass Lab of Università Politecnica delle Marche mainly during the application of a long-term monitoring of biomass quality employed in several installations. The woodchip samples comes from several power plants (district heating and combined heat and power plants) and different parts of Italy so that they could be considered representative of the national scene for number and type of suppliers, origin and biomass typology. It should be taken into account that the samples were chosen considering the requirements of UNI EN ISO 17225-4:2014 for the woodchip quality classes. As a consequence only samples with Ac at around 4% and GCV of at least 16,300 J/g were taken into account. According to UNI EN 14780:2011 standard, the material was first stabilized at 45 °C for 24 h then ground down to 1 mm of particle size by means of a cutting mill (mod. SM 2000, RETSCH). 2.2. Compositional analysis The analytical methodologies adopted for the determination of GCV and Ac refer to the standards UNI EN 14918:2010 and ISO 18122:2015, respectively. The ash content of air dried ground material (Acad) and its moisture content were determined using a thermo-gravimetric analyzer (mod.
78
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al.
3. Results and discussion
cross validation (RMSECV). Considering the huge number of samples, PLS model for the prediction of Ac was also validated using an external test set. The database was split; 109 samples were sent into the calibration set, and the remaining 16 samples were selected for the test set. The test set was chosen ad-hoc in order to maintain the representativeness of the training set. The figures of merit were taken into account in order to evaluate the performance of the model, in particular coefficient of determination (R2) and RMSECV. The ratio of standard error of prediction to standard deviation (RPD) and range to error ratio (RER) were also calculated. According to Williams [24], for screening application RPD value should be > 3, while RPD values > 5 can be used for quality control. RER values of < 6 indicate a poor model not recommended in any application; while values between 7 and 20 indicate the model as being fair and may be used in a screening application; values > 20 indicate good predictions suggesting the model could be used in any quality control application [12] [24,25]. It is well known that variables selection methods could enhance the performance of the prediction models in comparison to that of the fullspectrum PLS model [26]. Therefore, after the selection of the optimal number of LVs, the PLS models were calculated again on the wavelengths selected on the basis of different variable selections methods: regression vector (B), Variables Important in Projections (VIP) [21,27], Selectivity Ratio (SR) [28], interval Partial Least-Squares (iPLS) [29] and recursive weighted Partial Least Squares (rPLS) [30]. Variable selection methods were carried out both on the full spectrum (10,000 cm− 1 to 4000 cm− 1) and on the wavelength range between 7500 cm− 1 up to 4000 cm− 1, since the most significant information is contained in this range. All calculations included in this study – spectra pre-treatments, PLS modelling and variables selection – were performed in Matlab (ver. 7.10.0, The MathWorks) using in-house functions based on existing algorithms.
3.1. General statistical analysis Table 1 reports a general descriptive statistic of the two reference variables: Ac and GCV both air dried and on dry basis. The values were obtained from the calculation of all the samples both for GCV (86 samples) and Ac (125 samples). Pearson's correlation coefficient (r) was calculated. GCVad had no correlation with Acad (r = −0.112), while GCVdb had positive correlation with GCVad (r = 0.631) and Acdb had really high correlation with Acad (r = 0.999). Fig. 1 shows the box-plots of Ac and GCV both on air dried and on dry basis. Outliers can be noticed in GCVdb and GCVad plots. The statistical software used (Minitab Release 16) identifies as possible outlier the samples which fall outside the Q1–1.5 IQR or the Q3 + 1.5 IQR limits (shown on Fig. 1 as whiskers), where Q1 and Q3 are first and third quartiles, respectively, and IQR is the interquartile range. According to Tukey fences, a more conservative rule to mark probable outlier is used, that is Q1–3.0 IQR or the Q3 + 3.0 IQR. With this approach, only four samples in GCVdb were recognized as outliers. No outliers were removed before multivariate analysis. 3.2. Prediction of gross calorific value Residual vs. leverage, observed vs. predicted response and PLS score plots were taken into account and two outliers were found and deleted prior to the final PLS model. The same samples were recognized as probable outliers also according to Tukey fences in Fig. 1. Several preprocessing methods were evaluated both on GCVad and GCVdb (Table 2). Since the correlation between GCVad and GCVdb was 0.631 and the prediction results of GCVad was better than the one on a dry basis, further elaborations have taken into account only the GCVad. This also makes sense, as the NIR spectra were performed on the as-is samples. The strongest GCVad prediction model was developed using first derivative Savitzky-Golay with second-order polynomial and 21points window and average of the two replicates for each sample. The prediction model was developed using 5 LVs and had R2 = 0.81, RMSECV = 289 J/g, RPD = 2.5 and RER = 13.3 (Fig. 2.a) indicating the development of a good model that could be used as a screening tool for quality control application. This considering the fact that, according to UNI EN 14918:2010, the repeatability limit and reproducibility limit for GCV analysis are 140 J/g and 400 J/g respectively. The most important peaks in the prediction of GCVad are highlighted in fig. 2.b. The part of the spectra between 10,000 and 7500 cm− 1 doesn't contain significant information, in any case the full spectrum was used for PLS modelling.
Table 1 General descriptive statistics of woodchip samples.
Mean Standard deviation Min Max Range
Acad (%)
Acdb (%)
GCVad (J/g)
GCVdb (J/g)
2.20 0.98 0.38 4.14 3.76
2.30 1.03 0.40 4.36 3.96
18,376 709 16,330 20,186 3856
19,653 614 17,127 20,951 3824
Fig. 1. Box-plot for GCV (a) and Ac (b) values. The probable outliers were marked with circles according to Tukey fences.
79
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al.
Table 2 Summary of PLS prediction results for gross calorific value (J/g). The best model is highlighted in bold. Pre-treatments
Av SNV 13der1_av 21der1_av 13der2_av 21der2_av SNV_21der1_av
GCVad
GCVdb
R2
RMSECV
Bias
RPD
RER
LV
R2
RMSECV
Bias
RPD
RER
LV
0.71 0.66 0.80 0.81 0.76 0.75 0.75
355 385 299 289 323 328 328
−4.5 −1.9 −4.0 −3.8 −2.6 −7.4 −7.5
2.0 1.8 2.4 2.5 2.2 2.2 2.2
10.9 10.0 12.9 13.3 11.9 11.8 11.8
5 3 5 5 3 3 4
0.38 0.29 0.58 0.61 0.46 0.52 0.44
404 431 331 317 378 354 379
− 3.1 − 6.3 − 4.6 − 6.5 − 3.5 − 2.5 − 9.0
1.5 1.4 1.9 1.9 1.6 1.7 1.6
9.5 8.9 11.5 12.1 10.1 10.8 10.1
5 3 5 5 3 5 4
ad: air dried; db: dry basis; SNV: standard normal variate; Xder1: first derivative with X number of smoothing points; Xder2: second derivative with X number of smoothing points; av.: average; LV: number of latent variables.
Fig. 2. The final PLS model towards the gross calorific value was based on NIR data pre-processed by a first order derivation using the Savitzky-Golay algorithm with 21 smoothing points and a second order polynomial fitting. The PLS model uses 5 LVs. (a) regression plot of observed versus fitted response of gross calorific value air dried. (b) The corresponding PLS regression coefficients, with the important wavenumbers indicated by dotted vertical lines.
3.3. Prediction of ash content
The prediction of GCVad is mainly linked to CeH bonds containing more energy than CeC bonds and OeH bonds [14]. In any case, OeH bonds of H2O are also important since it is well know that GCV and moisture content are negatively correlated [13]. Band at 4154 and 4038 cm− 1 are related to the NIR region of CeH and CeC combinations. Peak at 4424 cm− 1 is characteristic to OeH and CeO stretching of acetyl groups [31], while peak at 4605 cm− 1 is assigned to CeH stretching and CeO stretching in lignin and hemicellulose [32]. Peaks at 5168 and 4705 cm− 1 are related to OeH combination. In particular, peak at 5168 cm− 1 is characteristic for the combination band of OeH stretching and OeH deformation vibration of water [31] and peak at 4705 cm− 1 for the OeH deformation and OeH stretching of lignin [32]. Finally, peak at 6025 cm− 1 corresponds to 1st OT CaromaticeH stretching.
Ac was estimated on both a wet and a dry basis. Different types of pre-processing methods were tested on the spectral data prior to PLS model building (Table 3). Since the correlation between Acad and Acdb was 0.999 and no difference were observed between the prediction results of the two models, further elaborations were on Acad only, as this is how the NIR instruments has seen the sample. Fagan et al. [12] reports similar observation in their work, too. Pre-processing does not seem to influence the prediction performance of the model, therefore only mean-centering was used. The prediction model was developed using 5 LVs and had R2 = 0.77, RMSECV = 0.47, RPD = 2.1 and RER = 8.0 (Fig. 3.a) indicating the potential use of the model for screening applications. This considering the fact that, according to ISO 18122:2015, the repeatability limit and reproducibility limit for Ac
80
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al.
Table 3 Summary of PLS prediction results for ash content (%). The best model is highlighted in bold. Pre-treatments
av SNV_av 13der1_av 13der2_av SNV_21der1_av
Ashad
Ashdb
R2
RMSECV
Bias
RPD
RER
LV
R2
RMSECV
Bias
RPD
RER
LV
0.77 0.76 0.76 0.74 0.75
0.47 0.48 0.48 0.50 0.49
− 0.002 − 0.002 − 0.006 − 0.002 0.006
2.1 2.0 2.0 2.0 2.0
8.0 7.8 7.8 7.5 7.7
5 4 2 4 4
0.77 0.77 0.76 0.75 0.76
0.49 0.49 0.50 0.51 0.50
− 0.012 − 0.010 − 0.011 − 0.003 0.006
2.1 2.1 2.1 2.0 2.1
8.1 8.1 7.9 7.8 7.9
5 4 3 4 4
ad: air dried; db: dry basis; SNV: standard normal variate; Xder1: first derivative with X number of smoothing points; Xder2: second derivative with X number of smoothing points; av.: average; LV: number of latent variables.
Fig. 3. The final PLS model towards the ash content was based on mean-centered NIR data. (a) regression plot of observed versus fitted response of ash content air dried. The PLS model uses 5 LVs. (b) PLS model regression coefficient plot with important wavenumbers indicated by dotted vertical lines.
stretching and OeH deformation vibration of water [31], while band at 4335 and 4007 cm− 1 is related to CeH stretching + CeC stretching [31].
analysis are: i) 0.1% and 0.2% absolute respectively if Ac is < 1%; ii) 10% and 20% relative if Ac is > 1%. Fig. 3.b reports the most relevant wavenumber in the prediction of ash content. The Ac prediction model is weaker than GCV model. Ash is the non-combustible portion of a biomass product [19] and it is difficult to predict since the inorganic constituent does not absorb in the near infrared region. However, inorganic compound can produce an infrared spectrum, but the infrared bands for inorganic materials are broader, fewer in number and appear at lower wavenumbers than those observed for organic materials [33]. Lestander and Rhen [14] reported that much of the inference relating to the prediction of Ac occurs in carbon bonds. In any case it is important to take into account that the prediction of Ac is an indirect prediction since it is based primarily on the presence of CeH and OeH bond containing compounds [13]. It is furthermore of interest to observe that the model only using meancentering is the optimal model, indicating that the scatter properties of the sample are related to its predicted ash content. Peaks at 6858, 6765, 5782, 4991 and 4748 cm− 1 are related to OeH bond, in particular peak at 6765 cm− 1 is assigned to 1st OT OeH stretching of cellulose, 5782 cm− 1 to 1st OT OeH of lignin and 4748 cm− 1 to OeH deformation + OeH stretching of cellulose [31]. Peak at 5199 cm− 1 is characteristic for the combination band of OeH
3.4. Variable selection With the aim to improve the performance of the models and give better predictions, different variable selection methods were performed both for Ac and GCV prediction models. PLS models on the reduced wavenumber range was compared to the full model using the figures of merit given in Table 2 and Table 3. VIP and SR did not give any improvement of the regression models. On the other side, better predictions were obtained using B, rPLS and iPLS. Table 4 reports the results of the two best variables selection methods obtained on the full FT-NIR spectral range (10,000–4000 cm− 1) and on a selected wavelength range (7500–4000 cm− 1). As it has already been noted in Fig. 2.b, the range 7500–4000 cm− 1 contains the most significant information for GCV prediction, therefore this part of the spectra was used in order to verify a possible improvement of the model. iPLS was tested using both forward and backwards methods and also different sizes of the intervals (15 and 31). It is important to note that the size of the intervals was kept constant in order to simplify the comparison. All the information are reported in Table 4.
81
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al. Table 4 Summary of variables selection results for Acad (%) and GCVad (J/g) prediction models. The best model is highlighted in bold. Parameter
Ac
GCV
Model
Spectral range
Type
n interval
PLS B rPLS (5LV) PLS B (5LV) rPLS (5LV) PLS iPLS for (10LV) rPLS (7LV) PLS iPLS for (10LV) rPLS (7LV)
– – – – – – – 50 – – 29 –
Optimal LVs
10,000–4000 10,000–4000 10,000–4000 7500–4000 7500–4000 7500–4000 10,000–4000 10,000–4000 10,000–4000 7500–4000 7500–4000 7500–4000
5 5 5 4 4 5 5 6 6 5 5 6
Bibliography wavenumber (cm− 1)
Assignment
4057–4069 4405–4420
4063 4411 4405 4404
CeH str. + CeC str. (C) [29] OeH str. + CeO str. (L) [29] OeH str. + CeO str. (C) [29] CeH2 str. + CeH2 deformation (C, H) [29] CeH str. + CeH deformation (H) [29] CeH + CeH combination CeH str. + C]O str. (L) [29] CareH str. + C]C str. (L, E) [29] CeH str. + C]O str. (H) [29] CH deformation and OH str. (C, L) [30] OeH asymmetric str. + OeH deformation of H2O (water) [29] 2nd OT C]O str. 1st OT CareH str. (E) () 1st OT CeH str. (H) [29] 1st OT OeH str. (L, E) [29]
4401 4489–4497 4563 4698–4709
4546 4686
4698–4709 4902–4921
4686 4904
5134–5168
5220–5150
5303–5311 5998–6025 7101–7104
5995 6003 7092
Prediction 2
RMSECV
R
0.47 0.50 0.44 0.47 0.45 0.44 289 258 247 286 237 235
0.77 0.74 0.80 0.77 0.78 0.80 0.81 0.85 0.86 0.82 0.87 0.87
RMSEP
R2
0.49 0.50 0.48 0.47 0.47 0.49 – – – – – –
0.78 0.75 0.77 0.80 0.79 0.77 – – – – – –
For Acad prediction, the range selection did not significantly improve the performance parameters of the PLS model, since some structured information is contained also after 7500 cm− 1 (see Fig. 3.b). Considering the full spectrum, the best model for Ac prediction was obtained using the rPLS method (5LVs). RMSECV value decreased from 0.47% to 0.44% and R2 increased from 0.77 to 0.80. Considering root mean square error of prediction (RMSEP), rPLS model are similar to the PLS model and the B model on the spectral range from 7500 to 4000 cm− 1. This indicates that the rPLS does not reduce the uncertainty as such, but it simplifies the model, thus making it easier to interpret. The variables selected by rPLS with the related band assignment were reported in Table 5 for GCVad and in Table 6 for Acad. Interpretation of FT-NIR spectra might be explained with support of fig. 4.a for GCVad and 4.b for Acad. First of all, it is important to note that not all the wavelengths already detected in the regression vector plots (fig. 2.b - 3.b) seem to be significant for the prediction of y-values. In comparison to our previous PLS models new wavenumbers have been selected by rPLS. In particular for GCVad prediction most of the new bands were linked to CeH bonds, i.e. wavelength range 4497–4489 cm− 1 and 4921–4902 cm− 1. Regarding the Acad, rPLS model selected some new variables in the first part of the spectra, in particular in the region between 7617 and 7563 cm− 1 corresponding to 1st OT of CeH combinations.
Table 5 infrared absorption band assignment associated with the variables selected by rPLS for the GCVad prediction. The wavelengths already detected as important for GCVad prediction and associated with high regression vector values were highlighted in bold (str.: stretching; OT: overtone; L: lignin; H: hemicellulose; C: cellulose; E: extractives). rPLS wavenumber range (cm− 1)
Cross validation
4. Conclusions Table 6 infrared absorption band assignment associated with the variables selected by rPLS for the Acad prediction. The wavelengths already detected as important for Acad prediction and associated with high regression vector values were highlighted in bold (str.: stretching; OT: overtone; L: lignin). rPLS wavenumber range (cm− 1)
Bibliography wavenumber (cm− 1)
4987 5188–5191
5220–5150
5793 6858 7563–7579 7617
5795
The results of this study demonstrated that coupling FT-NIR technique with multivariate analysis (Partial Least Squares) is a promising tool both for the prediction of GCV and Ac of woodchip samples. The performance of the model could be improved using rPLS; GCV was predicted with a RMSECV of 234 J/g and Ac with a RMSECV of 0.44%, making the model exploitable in screening applications. Moreover, the results also showed the potential to develop a robust FT-NIR predictive model to be used for industrial in-line applications in the woody biofuels sector, overcoming sampling problems and ensuring a total quality control. To take into account that the NIR method does not replace the laboratory methods, but it could be used in evaluation-screening process providing an indication of the biofuels quality. The great utility of such a method is based on the fact that, if applied in-line, all the incoming feedstock in a power plant could be checked providing a quality trend of the product. As a consequence, the technique represents an important decision-making tool for the different stakeholders involved in the wood chip energy chain. Moreover, the technique could also be used at-line or off-line. In this case, the main benefits coming from the implementation of this technology are several: fast execution, reduced costs and simplicity of analysis execution.
Assignment
OeH combination OeH asymmetric str. + OeH deformation of H2O (water) [29] 1st OT CeH str. (L) [29] 1st OT OeH str. 1st OT of CeH combinations 1st OT of CeH combinations
Comparing the different figures of merit, rPLS is the superior variable selection method. In particular, there is seen a significant improvement on the GCVad prediction, the model was obtained using rPLS method (7LVs) on the spectral range from 7500 to 4000 cm− 1. Here, the RMSECV value decreased from 289 J/g to 234 J/g and R2 increased from 0.81 to 0.87. 82
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al.
Fig. 4. (a) variables selected by rPLS for GCVad prediction. (b) variables selected by rPLS for Acad prediction. The dotted vertical lines indicate the wavenumbers indicated as important in the PLS model regression coefficient plot (Fig. 2.b and 3.b). (2012) 203–211. [16] C. Huang, et al., Ultimate analysis and heating value prediction of straw by near infrared spectroscopy, Waste Manag. 29 (6) (2009) 1793–1797. [17] U.K. Saha, L. Sonon, M. Kane, Prediction of calorific values, moisture, ash, carbon, nitrogen, and sulfur content of pine tree biomass using near infrared spectroscopy, J. Near Infrared Spectrosc. 25 (4) (2017) 242–255. [18] J. Posom, P. Sirisomboon, Evaluation of lower heating value and elemental composition of bamboo using near infrared spectroscopy, Energy 121 (Supplement C) (2017) 147–158. [19] S. Bruun, et al., Prediction of the degradability and ash content of wheat straw from different cultivars using near infrared spectroscopy, Ind. Crop. Prod. 31 (2) (2010) 321–326. [20] M. Blanco, I. Villarroya, NIR spectroscopy: a rapid-response analytical tool, TrAC Trends Anal. Chem. 21 (4) (2002) 240–250. [21] S. Wold, M. Sjöström, L. Eriksson, PLS-regression: a basic tool of chemometrics, Chemom. Intell. Lab. Syst. 58 (2) (2001) 109–130. [22] Å. Rinnan, F.V.D. Berg, S.B. Engelsen, Review of the most common pre-processing techniques for near-infrared spectra, TrAC Trends Anal. Chem. 28 (10) (2009) 1201–1222. [23] A. Savitzky, M.J.E. Golay, Smoothing and differentiation of data by simplified least squares procedures, Anal. Chem. 36 (8) (1964) 1627–1639. [24] P.C. Williams, Implementation of near-infrared technology, in: P.C. Williams, K. Norris (Eds.), Near Infrared Technology in the Agriculture and Food Industries, 2nd ed, American Association of Cereal Chemists, Minnesota, USA, 2001. [25] C.C. Fagan, et al., Evaluating Mid-infrared spectroscopy as a new technique for predicting sensory texture attributes of processed cheese, J. Dairy Sci. 90 (3) (2007) 1122–1132. [26] C.M. Andersen, R. Bro, Variable selection in regression—a tutorial, J. Chemom. 24 (11 − 12) (2010) 728–737. [27] I.-G. Chong, C.-H. Jun, Performance of some variable selection methods when multicollinearity is present, Chemom. Intell. Lab. Syst. 78 (1–2) (2005) 103–112. [28] T. Rajalahti, et al., Biomarker discovery in mass spectral profiles by means of selectivity ratio plot, Chemom. Intell. Lab. Syst. 95 (1) (2009) 35–48. [29] L. Norgaard, et al., Interval Partial Least-Squares Regression (iPLS): a comparative chemometric study with an example from near-infrared spectroscopy, Appl. Spectrosc. 54 (3) (2000) 413–419. [30] Å. Rinnan, et al., Recursive weighted partial least squares (rPLS): an efficient variable selection method using PLS, J. Chemom. 28 (5) (2014) 439–447. [31] M. Schwanninger, J. Rodrigues, K. Fackler, A review of band assignments in near infrared spectra of wood and wood components, J. Near Infrared Spectrosc. 19 (5) (2011) 287–308. [32] C.-M. Popescu, M.-C. Popescu, A near infrared spectroscopic study of the structural modifications of lime (Tilia cordata Mill.) wood during hydro-thermal treatment, Spectrochim. Acta A Mol. Biomol. Spectrosc. 115 (2013) 227–233. [33] B.H. Stuart, Inorganic molecules, Infrared Spectroscopy: Fundamentals and Applications, 95-111 John Wiley & Sons, Ltd., 2005.
Acknowledgements The authors would like to thank Thermo Fisher Scientific, for the support in the development of this study. References [1] A. González, et al., Review of micro- and small-scale technologies to produce electricity and heat from Mediterranean forests' wood chips, Renew. Sust. Energ. Rev. 43 (0) (2015) 143–155. [2] M. Manzone, Quality, productivity, energy and costs of woodchip produced by Cedrus deodara plantations: a case study in Italy, Biomass Bioenergy 92 (2016) 81–87. [3] D.Y. Guimier, M.A. Pottie, Preparation of Forest Biomass for Optimal Conversion, Forest Engineering Research Institute of Canada, 1985. [4] G. Toscano, et al., Investigation of woodchip quality: relationship between the most important chemical and physical parameters, Energy 106 (2016) 38–44. [5] P. Dare, et al., Combustion performance of biomass residue and purpose grown species, Biomass Bioenergy 21 (4) (2001) 277–287. [6] B.R. Hames, et al., Rapid biomass analysis: New tools for compositional analysis of corn stover feedstocks and process intermediates from ethanol production, Appl. Biochem. Biotechnol. Part A Enzyme Eng. Biotechnol. 105 (1–3) (2003) 5–16. [7] S.V. Vassilev, et al., An overview of the chemical composition of biomass, Fuel 89 (5) (2010) 913–933. [8] 17225-4, E.I., solid biofuels - fuel specifications and classes - part 4, Graded Wood Chips, 2014. [9] J. Posom, et al., Rapid non-destructive evaluation of moisture content and higher heating value of Leucaena leucocephala pellets using near infrared spectroscopy, Energy 107 (2016) 464–472. [10] I. Lewandowski, A. Kicherer, Combustion quality of biomass: practical relevance and experiments to modify the biomass quality of Miscanthus × giganteus, Eur. J. Agron. 6 (3–4) (1997) 163–177. [11] M.K. Misra, K.W. Ragland, A.J. Baker, Wood ash composition as a function of furnace temperature, Biomass Bioenergy 4 (2) (1993) 103–116. [12] C.C. Fagan, C.D. Everard, K. McDonnell, Prediction of moisture, calorific value, ash and carbon content of two dedicated bioenergy crops using near-infrared spectroscopy, Bioresour. Technol. 102 (8) (2011) 5200–5206. [13] Gillespie, G.D., C.D. Everard, and K.P. McDonnell, Prediction of biomass pellet quality indices using near infrared spectroscopy. Energy, (0). [14] T.A. Lestander, C. Rhen, Multivariate NIR spectroscopy models for moisture, ash and calorific content in biofuels using bi-orthogonal partial least squares regression, Analyst 130 (8) (2005) 1182–1189. [15] C.D. Everard, K.P. McDonnell, C.C. Fagan, Prediction of biomass gross calorific values using visible and near infrared spectroscopy, Biomass Bioenergy 45 (0)
83
Contents lists available at ScienceDirect
Fuel Processing Technology journal homepage: www.elsevier.com/locate/fuproc
Prediction of gross calorific value and ash content of woodchip samples by means of FT-NIR spectroscopy
MARK
M. Mancinia, Å. Rinnanb, A. Pizzia, G. Toscanoa,⁎ a b
Agricultural, Food and Environmental Sciences, Università Politecnica delle Marche, via Brecce Bianche, 60131 Ancona, Italy Department of Food Science, Faculty of Science, University of Copenhagen, Rolighedsvej 26, 1958 Frederiksberg C, Denmark
A R T I C L E I N F O
A B S T R A C T
Keywords: PLS Variables selection rPLS Biomass quality Woodchip
The use of woodchip, and biofuels in general, is a fundamental step towards the transition from fossil fuel to renewable energy. The growth in the demand for wood fuels and the inherent variability in the properties of woody material lead to the need to verify its quality. EN ISO 17225-4 divides woodchip in different quality classes according to chemical-physical parameters and quality attributes. In this study, we have coupled near infrared spectroscopy with Partial Least Square regression to model gross calorific value and ash content of woodchip samples. Moreover, variables selection methods were tested in order to improve the models and get better prediction. Gross calorific value and ash content were predicted with a standard error of 234 J/g and 0.44%, respectively. The results indicate that the models could be used in screening applications and near infrared spectroscopy is a promising tool in the evaluation of biomass quality.
1. Introduction In recent years, the interest in bioenergy as an alternative to fossil energy is increasing because of the continuously growing energy demand, the decreasing availability of fossil fuels and the need for a reduction of environmental impact. In order to reach these targets European policies is aiming to promote the use of renewable energy sources. Woody biomass is such a source of energy, it is present more or less everywhere, is available in many forms and can be easily stored, especially in comparison with other energy sources [1]. In particular, woodchip is really appreciated because it consists of homogeneous particles with a specific size and it guarantees benefits in terms of increased load density and handling quality [2,3]. In different European countries, and in Italy as well, the number of power plants fueled with woodchip is increasing and accordingly also the demand for wood fuels. As a consequence, the biomass quality could experience a decrease and need to be analysed [4]. Moreover, it is known that there is an inherent variability in the properties of woody material that is influenced by many factors [5,6] and this leads to the need to employ quality standards in order to check the quality of the product. CEN/TC 335 has established a number of standards to ensure biomass quality. Chemical-physical parameters and quality features (e.g. origin and source) divide the biomass in different quality classes. The
⁎
identification and characterization of chemical composition of a solid fuel is the most important step during the investigation and application of such fuel. This composition is a unique code that characterizes and determines the properties, quality, application perspectives and environmental problems related to any fuel [7]. EN ISO 17225–4 [8] is the quality standard for graded woodchip for residential, small commercial and public building applications and provides limits for three different woodchip quality classes: A1, A2 and B. In particular, calorific value is an essential parameter in the specification of biomass quality to set the price of the product. Moreover it is an important characteristic for the planning and control of power plants using biomass fuel [9]. Ash content (Ac) influences combustion efficiency and may cause cleaning and combustion problems, such as slagging in furnaces, fouling of heat exchanger surfaces, corrosion in the combustion device [4,10,11]. Ac is also the most important discriminant parameter among the woodchip quality classes that are related to a maximum Ac of 1.0, 1.5 and 3.0% respectively. Gross calorific value (GCV) and Ac are normally determined by traditional laboratory analysis, but the process is tedious, expensive and requires specialized experts. As a consequence, it is necessary to develop a technique that is rapid, economic and simple. A good candidate could be Near Infrared (NIR) spectroscopy which is already widely used for quantitative and qualitative purposes in different sectors, i.e. pharmacy, food and agricultural industries.
Corresponding author. E-mail address: [email protected] (G. Toscano).
http://dx.doi.org/10.1016/j.fuproc.2017.09.021 Received 23 May 2017; Received in revised form 19 September 2017; Accepted 20 September 2017 0378-3820/ © 2017 Elsevier B.V. All rights reserved.
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al.
701 Leco). The Acad was determined as the ratio between the residue remaining after the sample was heated in air at 550 ± 10 °C to initial biomass. The average Acad was calculated based on two measurements per sample. The ash content on a dry basis (Acdb) was obtained for each sample by multiplying Acad by its dry matter content in percentage. A subset of samples (n = 86) was selected for GCV analysis. Gross calorific value of air dried ground material (GCVad) was determined in a dynamic mode at 25 °C using a bomb calorimeter (mod.C2000 basic, IKA). The calorimeter was calibrated using a benzoic acid standard (IKA Benzoic Acid C723). The analyses were performed in duplicate for each sample. As for Ac, gross calorific value on a dry basis (GCVdb) was obtained for each sample by multiplying GCVad by the dry matter content in percentage.
A literature research, shows that only a limited number of studies has been performed on the use of NIR spectroscopy in the field of biomass control. Fagan et al. [12] and Gillespie et al. [13] examined the possibility to employ NIR spectroscopy and chemometrics to predict moisture, ash, carbon contents and gross calorific value of Miscanthus, willow and herbaceous energy grasses. The possibility to predict calorific value, moisture and ash content was studied also by Lestander and Rhen [14] on Norway spruce (Picea abies (L.) Karst.). Moreover, studies have shown the potential of NIR spectroscopy to predict calorific value in dedicated bioenergy crops [15], in Leucaena leucocephala pellets [9] and in straw [16]. Saha et al. [17] investigated the possibility to predict calorific values, moisture, ash, carbon, nitrogen, and sulfur content of pine tree biomass using near infrared spectroscopy. Lower heating value and elemental composition were also examined by Pasom et al. using partial least squares regression and considering bamboo samples [18]. The ash content was also investigated in wheat straw by Bruun et al. with good results [19]. All the prediction models developed in the aforementioned studies were based on a specific kind of biomass or wood species. To the authors' knowledge, there are no works based on woodchip samples taken directly from the market, without having information about wood typology. This approach allows the development of more robust and reliable prediction models, which could be employed on a national scale. Furthermore, in the solid biofuels sector, no variable selection methods were applied to NIR models for the improvements of the models performance. The objective of this study is to develop models based on Fourier Transform Infrared (FT-NIR) Spectroscopy and multivariate analysis (Partial Least Squares) for the prediction of GCV and Ac of woodchip samples. In order to achieve this, a large number of samples from several power plants, representative of the national scene, were prepared and analysed by FT-NIR spectroscopy. The multivariate regression models were further elaborated using chemometric techniques, such as variables selection combined with data pre-processing, in order to get better prediction models.
2.3. Near-infrared spectroscopy The near infrared spectra were collected using a FT-NIR spectrophotometer (mod. Antaris II, Thermo Fisher Scientific Inc., USA) equipped with a halogen lamp as a source and an InGaAs detector. The samples were acquired in diffuse reflectance mode using an integrating sphere and were kept in rotation during the acquisition by means of a sample cup spinner to increase the representativeness of the material. Each spectrum has been computed as an average of 32 successive scans acquired at a wavelength range from 10,000 to 4000 cm− 1 and with the spectral resolution of 8 cm− 1. All measurements have been performed at room temperature (18–20 °C) and each sample is recorded in duplicate. A blank spectrum was acquired every hour to exclude the signals not associated to the sample, but to the instrument or environment. 2.4. Data processing and Partial Least Square Regression NIR spectroscopy requires chemometrics to extract as much relevant information as possible from the analytical data [20]. Partial Least Square Regression (PLS) is a common method applied in spectroscopy for quantitative analysis. It finds the relationship between a y-value the parameter to be quantified - and the spectral data matrix, maximizing the covariation between them. PLS finds a new smaller set of variables, called latent variables (LVs), that are linear combinations of the spectral data and relevant for the determination of the parameters of interest [21]. In this study, PLS models were calculated to predict Ac and GCV of woodchip samples both air dried and on the dry basis. Prior to PLS regression, in order to minimize the effect of baseline shifts and noise, the spectra were pre-processed [22]. Different pre-treatments including Standard Normal Variate (SNV), first and second derivative spectra (Savitzky-Golay [23] with 13 or 21 smoothing points and 2nd order polynomial) were applied. As no replicate outliers were found, each sample was averaged across the replicates. Residual vs. leverage, observed vs. predicted response and PLS score plots were utilized in order to identify possible sample outliers. PLS regression models were validated using Venetian blind-cross validation (5 segments). In this type of validation, n validation models are created from the original data set in order to assess the performance of the prediction model. For each validation model, the test set is obtained taking out samples at every n position in the original data set, while the remaining samples represent the training set which is used to build up the model. Therefore, a group of samples is left out from the calibration data set and the model is calibrated on the remaining data point. Then the values for the test set are predicted and the prediction residuals are calculated. The procedure is repeated until each subset has been left out once. As a result, all prediction residuals are combined to calculate validation residual variance and the root mean square error of
2. Materials and methods 2.1. Sample collection and preparation In this study a total of 125 woodchip samples were collected and analysed by the Biomass Lab of Università Politecnica delle Marche mainly during the application of a long-term monitoring of biomass quality employed in several installations. The woodchip samples comes from several power plants (district heating and combined heat and power plants) and different parts of Italy so that they could be considered representative of the national scene for number and type of suppliers, origin and biomass typology. It should be taken into account that the samples were chosen considering the requirements of UNI EN ISO 17225-4:2014 for the woodchip quality classes. As a consequence only samples with Ac at around 4% and GCV of at least 16,300 J/g were taken into account. According to UNI EN 14780:2011 standard, the material was first stabilized at 45 °C for 24 h then ground down to 1 mm of particle size by means of a cutting mill (mod. SM 2000, RETSCH). 2.2. Compositional analysis The analytical methodologies adopted for the determination of GCV and Ac refer to the standards UNI EN 14918:2010 and ISO 18122:2015, respectively. The ash content of air dried ground material (Acad) and its moisture content were determined using a thermo-gravimetric analyzer (mod.
78
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al.
3. Results and discussion
cross validation (RMSECV). Considering the huge number of samples, PLS model for the prediction of Ac was also validated using an external test set. The database was split; 109 samples were sent into the calibration set, and the remaining 16 samples were selected for the test set. The test set was chosen ad-hoc in order to maintain the representativeness of the training set. The figures of merit were taken into account in order to evaluate the performance of the model, in particular coefficient of determination (R2) and RMSECV. The ratio of standard error of prediction to standard deviation (RPD) and range to error ratio (RER) were also calculated. According to Williams [24], for screening application RPD value should be > 3, while RPD values > 5 can be used for quality control. RER values of < 6 indicate a poor model not recommended in any application; while values between 7 and 20 indicate the model as being fair and may be used in a screening application; values > 20 indicate good predictions suggesting the model could be used in any quality control application [12] [24,25]. It is well known that variables selection methods could enhance the performance of the prediction models in comparison to that of the fullspectrum PLS model [26]. Therefore, after the selection of the optimal number of LVs, the PLS models were calculated again on the wavelengths selected on the basis of different variable selections methods: regression vector (B), Variables Important in Projections (VIP) [21,27], Selectivity Ratio (SR) [28], interval Partial Least-Squares (iPLS) [29] and recursive weighted Partial Least Squares (rPLS) [30]. Variable selection methods were carried out both on the full spectrum (10,000 cm− 1 to 4000 cm− 1) and on the wavelength range between 7500 cm− 1 up to 4000 cm− 1, since the most significant information is contained in this range. All calculations included in this study – spectra pre-treatments, PLS modelling and variables selection – were performed in Matlab (ver. 7.10.0, The MathWorks) using in-house functions based on existing algorithms.
3.1. General statistical analysis Table 1 reports a general descriptive statistic of the two reference variables: Ac and GCV both air dried and on dry basis. The values were obtained from the calculation of all the samples both for GCV (86 samples) and Ac (125 samples). Pearson's correlation coefficient (r) was calculated. GCVad had no correlation with Acad (r = −0.112), while GCVdb had positive correlation with GCVad (r = 0.631) and Acdb had really high correlation with Acad (r = 0.999). Fig. 1 shows the box-plots of Ac and GCV both on air dried and on dry basis. Outliers can be noticed in GCVdb and GCVad plots. The statistical software used (Minitab Release 16) identifies as possible outlier the samples which fall outside the Q1–1.5 IQR or the Q3 + 1.5 IQR limits (shown on Fig. 1 as whiskers), where Q1 and Q3 are first and third quartiles, respectively, and IQR is the interquartile range. According to Tukey fences, a more conservative rule to mark probable outlier is used, that is Q1–3.0 IQR or the Q3 + 3.0 IQR. With this approach, only four samples in GCVdb were recognized as outliers. No outliers were removed before multivariate analysis. 3.2. Prediction of gross calorific value Residual vs. leverage, observed vs. predicted response and PLS score plots were taken into account and two outliers were found and deleted prior to the final PLS model. The same samples were recognized as probable outliers also according to Tukey fences in Fig. 1. Several preprocessing methods were evaluated both on GCVad and GCVdb (Table 2). Since the correlation between GCVad and GCVdb was 0.631 and the prediction results of GCVad was better than the one on a dry basis, further elaborations have taken into account only the GCVad. This also makes sense, as the NIR spectra were performed on the as-is samples. The strongest GCVad prediction model was developed using first derivative Savitzky-Golay with second-order polynomial and 21points window and average of the two replicates for each sample. The prediction model was developed using 5 LVs and had R2 = 0.81, RMSECV = 289 J/g, RPD = 2.5 and RER = 13.3 (Fig. 2.a) indicating the development of a good model that could be used as a screening tool for quality control application. This considering the fact that, according to UNI EN 14918:2010, the repeatability limit and reproducibility limit for GCV analysis are 140 J/g and 400 J/g respectively. The most important peaks in the prediction of GCVad are highlighted in fig. 2.b. The part of the spectra between 10,000 and 7500 cm− 1 doesn't contain significant information, in any case the full spectrum was used for PLS modelling.
Table 1 General descriptive statistics of woodchip samples.
Mean Standard deviation Min Max Range
Acad (%)
Acdb (%)
GCVad (J/g)
GCVdb (J/g)
2.20 0.98 0.38 4.14 3.76
2.30 1.03 0.40 4.36 3.96
18,376 709 16,330 20,186 3856
19,653 614 17,127 20,951 3824
Fig. 1. Box-plot for GCV (a) and Ac (b) values. The probable outliers were marked with circles according to Tukey fences.
79
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al.
Table 2 Summary of PLS prediction results for gross calorific value (J/g). The best model is highlighted in bold. Pre-treatments
Av SNV 13der1_av 21der1_av 13der2_av 21der2_av SNV_21der1_av
GCVad
GCVdb
R2
RMSECV
Bias
RPD
RER
LV
R2
RMSECV
Bias
RPD
RER
LV
0.71 0.66 0.80 0.81 0.76 0.75 0.75
355 385 299 289 323 328 328
−4.5 −1.9 −4.0 −3.8 −2.6 −7.4 −7.5
2.0 1.8 2.4 2.5 2.2 2.2 2.2
10.9 10.0 12.9 13.3 11.9 11.8 11.8
5 3 5 5 3 3 4
0.38 0.29 0.58 0.61 0.46 0.52 0.44
404 431 331 317 378 354 379
− 3.1 − 6.3 − 4.6 − 6.5 − 3.5 − 2.5 − 9.0
1.5 1.4 1.9 1.9 1.6 1.7 1.6
9.5 8.9 11.5 12.1 10.1 10.8 10.1
5 3 5 5 3 5 4
ad: air dried; db: dry basis; SNV: standard normal variate; Xder1: first derivative with X number of smoothing points; Xder2: second derivative with X number of smoothing points; av.: average; LV: number of latent variables.
Fig. 2. The final PLS model towards the gross calorific value was based on NIR data pre-processed by a first order derivation using the Savitzky-Golay algorithm with 21 smoothing points and a second order polynomial fitting. The PLS model uses 5 LVs. (a) regression plot of observed versus fitted response of gross calorific value air dried. (b) The corresponding PLS regression coefficients, with the important wavenumbers indicated by dotted vertical lines.
3.3. Prediction of ash content
The prediction of GCVad is mainly linked to CeH bonds containing more energy than CeC bonds and OeH bonds [14]. In any case, OeH bonds of H2O are also important since it is well know that GCV and moisture content are negatively correlated [13]. Band at 4154 and 4038 cm− 1 are related to the NIR region of CeH and CeC combinations. Peak at 4424 cm− 1 is characteristic to OeH and CeO stretching of acetyl groups [31], while peak at 4605 cm− 1 is assigned to CeH stretching and CeO stretching in lignin and hemicellulose [32]. Peaks at 5168 and 4705 cm− 1 are related to OeH combination. In particular, peak at 5168 cm− 1 is characteristic for the combination band of OeH stretching and OeH deformation vibration of water [31] and peak at 4705 cm− 1 for the OeH deformation and OeH stretching of lignin [32]. Finally, peak at 6025 cm− 1 corresponds to 1st OT CaromaticeH stretching.
Ac was estimated on both a wet and a dry basis. Different types of pre-processing methods were tested on the spectral data prior to PLS model building (Table 3). Since the correlation between Acad and Acdb was 0.999 and no difference were observed between the prediction results of the two models, further elaborations were on Acad only, as this is how the NIR instruments has seen the sample. Fagan et al. [12] reports similar observation in their work, too. Pre-processing does not seem to influence the prediction performance of the model, therefore only mean-centering was used. The prediction model was developed using 5 LVs and had R2 = 0.77, RMSECV = 0.47, RPD = 2.1 and RER = 8.0 (Fig. 3.a) indicating the potential use of the model for screening applications. This considering the fact that, according to ISO 18122:2015, the repeatability limit and reproducibility limit for Ac
80
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al.
Table 3 Summary of PLS prediction results for ash content (%). The best model is highlighted in bold. Pre-treatments
av SNV_av 13der1_av 13der2_av SNV_21der1_av
Ashad
Ashdb
R2
RMSECV
Bias
RPD
RER
LV
R2
RMSECV
Bias
RPD
RER
LV
0.77 0.76 0.76 0.74 0.75
0.47 0.48 0.48 0.50 0.49
− 0.002 − 0.002 − 0.006 − 0.002 0.006
2.1 2.0 2.0 2.0 2.0
8.0 7.8 7.8 7.5 7.7
5 4 2 4 4
0.77 0.77 0.76 0.75 0.76
0.49 0.49 0.50 0.51 0.50
− 0.012 − 0.010 − 0.011 − 0.003 0.006
2.1 2.1 2.1 2.0 2.1
8.1 8.1 7.9 7.8 7.9
5 4 3 4 4
ad: air dried; db: dry basis; SNV: standard normal variate; Xder1: first derivative with X number of smoothing points; Xder2: second derivative with X number of smoothing points; av.: average; LV: number of latent variables.
Fig. 3. The final PLS model towards the ash content was based on mean-centered NIR data. (a) regression plot of observed versus fitted response of ash content air dried. The PLS model uses 5 LVs. (b) PLS model regression coefficient plot with important wavenumbers indicated by dotted vertical lines.
stretching and OeH deformation vibration of water [31], while band at 4335 and 4007 cm− 1 is related to CeH stretching + CeC stretching [31].
analysis are: i) 0.1% and 0.2% absolute respectively if Ac is < 1%; ii) 10% and 20% relative if Ac is > 1%. Fig. 3.b reports the most relevant wavenumber in the prediction of ash content. The Ac prediction model is weaker than GCV model. Ash is the non-combustible portion of a biomass product [19] and it is difficult to predict since the inorganic constituent does not absorb in the near infrared region. However, inorganic compound can produce an infrared spectrum, but the infrared bands for inorganic materials are broader, fewer in number and appear at lower wavenumbers than those observed for organic materials [33]. Lestander and Rhen [14] reported that much of the inference relating to the prediction of Ac occurs in carbon bonds. In any case it is important to take into account that the prediction of Ac is an indirect prediction since it is based primarily on the presence of CeH and OeH bond containing compounds [13]. It is furthermore of interest to observe that the model only using meancentering is the optimal model, indicating that the scatter properties of the sample are related to its predicted ash content. Peaks at 6858, 6765, 5782, 4991 and 4748 cm− 1 are related to OeH bond, in particular peak at 6765 cm− 1 is assigned to 1st OT OeH stretching of cellulose, 5782 cm− 1 to 1st OT OeH of lignin and 4748 cm− 1 to OeH deformation + OeH stretching of cellulose [31]. Peak at 5199 cm− 1 is characteristic for the combination band of OeH
3.4. Variable selection With the aim to improve the performance of the models and give better predictions, different variable selection methods were performed both for Ac and GCV prediction models. PLS models on the reduced wavenumber range was compared to the full model using the figures of merit given in Table 2 and Table 3. VIP and SR did not give any improvement of the regression models. On the other side, better predictions were obtained using B, rPLS and iPLS. Table 4 reports the results of the two best variables selection methods obtained on the full FT-NIR spectral range (10,000–4000 cm− 1) and on a selected wavelength range (7500–4000 cm− 1). As it has already been noted in Fig. 2.b, the range 7500–4000 cm− 1 contains the most significant information for GCV prediction, therefore this part of the spectra was used in order to verify a possible improvement of the model. iPLS was tested using both forward and backwards methods and also different sizes of the intervals (15 and 31). It is important to note that the size of the intervals was kept constant in order to simplify the comparison. All the information are reported in Table 4.
81
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al. Table 4 Summary of variables selection results for Acad (%) and GCVad (J/g) prediction models. The best model is highlighted in bold. Parameter
Ac
GCV
Model
Spectral range
Type
n interval
PLS B rPLS (5LV) PLS B (5LV) rPLS (5LV) PLS iPLS for (10LV) rPLS (7LV) PLS iPLS for (10LV) rPLS (7LV)
– – – – – – – 50 – – 29 –
Optimal LVs
10,000–4000 10,000–4000 10,000–4000 7500–4000 7500–4000 7500–4000 10,000–4000 10,000–4000 10,000–4000 7500–4000 7500–4000 7500–4000
5 5 5 4 4 5 5 6 6 5 5 6
Bibliography wavenumber (cm− 1)
Assignment
4057–4069 4405–4420
4063 4411 4405 4404
CeH str. + CeC str. (C) [29] OeH str. + CeO str. (L) [29] OeH str. + CeO str. (C) [29] CeH2 str. + CeH2 deformation (C, H) [29] CeH str. + CeH deformation (H) [29] CeH + CeH combination CeH str. + C]O str. (L) [29] CareH str. + C]C str. (L, E) [29] CeH str. + C]O str. (H) [29] CH deformation and OH str. (C, L) [30] OeH asymmetric str. + OeH deformation of H2O (water) [29] 2nd OT C]O str. 1st OT CareH str. (E) () 1st OT CeH str. (H) [29] 1st OT OeH str. (L, E) [29]
4401 4489–4497 4563 4698–4709
4546 4686
4698–4709 4902–4921
4686 4904
5134–5168
5220–5150
5303–5311 5998–6025 7101–7104
5995 6003 7092
Prediction 2
RMSECV
R
0.47 0.50 0.44 0.47 0.45 0.44 289 258 247 286 237 235
0.77 0.74 0.80 0.77 0.78 0.80 0.81 0.85 0.86 0.82 0.87 0.87
RMSEP
R2
0.49 0.50 0.48 0.47 0.47 0.49 – – – – – –
0.78 0.75 0.77 0.80 0.79 0.77 – – – – – –
For Acad prediction, the range selection did not significantly improve the performance parameters of the PLS model, since some structured information is contained also after 7500 cm− 1 (see Fig. 3.b). Considering the full spectrum, the best model for Ac prediction was obtained using the rPLS method (5LVs). RMSECV value decreased from 0.47% to 0.44% and R2 increased from 0.77 to 0.80. Considering root mean square error of prediction (RMSEP), rPLS model are similar to the PLS model and the B model on the spectral range from 7500 to 4000 cm− 1. This indicates that the rPLS does not reduce the uncertainty as such, but it simplifies the model, thus making it easier to interpret. The variables selected by rPLS with the related band assignment were reported in Table 5 for GCVad and in Table 6 for Acad. Interpretation of FT-NIR spectra might be explained with support of fig. 4.a for GCVad and 4.b for Acad. First of all, it is important to note that not all the wavelengths already detected in the regression vector plots (fig. 2.b - 3.b) seem to be significant for the prediction of y-values. In comparison to our previous PLS models new wavenumbers have been selected by rPLS. In particular for GCVad prediction most of the new bands were linked to CeH bonds, i.e. wavelength range 4497–4489 cm− 1 and 4921–4902 cm− 1. Regarding the Acad, rPLS model selected some new variables in the first part of the spectra, in particular in the region between 7617 and 7563 cm− 1 corresponding to 1st OT of CeH combinations.
Table 5 infrared absorption band assignment associated with the variables selected by rPLS for the GCVad prediction. The wavelengths already detected as important for GCVad prediction and associated with high regression vector values were highlighted in bold (str.: stretching; OT: overtone; L: lignin; H: hemicellulose; C: cellulose; E: extractives). rPLS wavenumber range (cm− 1)
Cross validation
4. Conclusions Table 6 infrared absorption band assignment associated with the variables selected by rPLS for the Acad prediction. The wavelengths already detected as important for Acad prediction and associated with high regression vector values were highlighted in bold (str.: stretching; OT: overtone; L: lignin). rPLS wavenumber range (cm− 1)
Bibliography wavenumber (cm− 1)
4987 5188–5191
5220–5150
5793 6858 7563–7579 7617
5795
The results of this study demonstrated that coupling FT-NIR technique with multivariate analysis (Partial Least Squares) is a promising tool both for the prediction of GCV and Ac of woodchip samples. The performance of the model could be improved using rPLS; GCV was predicted with a RMSECV of 234 J/g and Ac with a RMSECV of 0.44%, making the model exploitable in screening applications. Moreover, the results also showed the potential to develop a robust FT-NIR predictive model to be used for industrial in-line applications in the woody biofuels sector, overcoming sampling problems and ensuring a total quality control. To take into account that the NIR method does not replace the laboratory methods, but it could be used in evaluation-screening process providing an indication of the biofuels quality. The great utility of such a method is based on the fact that, if applied in-line, all the incoming feedstock in a power plant could be checked providing a quality trend of the product. As a consequence, the technique represents an important decision-making tool for the different stakeholders involved in the wood chip energy chain. Moreover, the technique could also be used at-line or off-line. In this case, the main benefits coming from the implementation of this technology are several: fast execution, reduced costs and simplicity of analysis execution.
Assignment
OeH combination OeH asymmetric str. + OeH deformation of H2O (water) [29] 1st OT CeH str. (L) [29] 1st OT OeH str. 1st OT of CeH combinations 1st OT of CeH combinations
Comparing the different figures of merit, rPLS is the superior variable selection method. In particular, there is seen a significant improvement on the GCVad prediction, the model was obtained using rPLS method (7LVs) on the spectral range from 7500 to 4000 cm− 1. Here, the RMSECV value decreased from 289 J/g to 234 J/g and R2 increased from 0.81 to 0.87. 82
Fuel Processing Technology 169 (2018) 77–83
M. Mancini et al.
Fig. 4. (a) variables selected by rPLS for GCVad prediction. (b) variables selected by rPLS for Acad prediction. The dotted vertical lines indicate the wavenumbers indicated as important in the PLS model regression coefficient plot (Fig. 2.b and 3.b). (2012) 203–211. [16] C. Huang, et al., Ultimate analysis and heating value prediction of straw by near infrared spectroscopy, Waste Manag. 29 (6) (2009) 1793–1797. [17] U.K. Saha, L. Sonon, M. Kane, Prediction of calorific values, moisture, ash, carbon, nitrogen, and sulfur content of pine tree biomass using near infrared spectroscopy, J. Near Infrared Spectrosc. 25 (4) (2017) 242–255. [18] J. Posom, P. Sirisomboon, Evaluation of lower heating value and elemental composition of bamboo using near infrared spectroscopy, Energy 121 (Supplement C) (2017) 147–158. [19] S. Bruun, et al., Prediction of the degradability and ash content of wheat straw from different cultivars using near infrared spectroscopy, Ind. Crop. Prod. 31 (2) (2010) 321–326. [20] M. Blanco, I. Villarroya, NIR spectroscopy: a rapid-response analytical tool, TrAC Trends Anal. Chem. 21 (4) (2002) 240–250. [21] S. Wold, M. Sjöström, L. Eriksson, PLS-regression: a basic tool of chemometrics, Chemom. Intell. Lab. Syst. 58 (2) (2001) 109–130. [22] Å. Rinnan, F.V.D. Berg, S.B. Engelsen, Review of the most common pre-processing techniques for near-infrared spectra, TrAC Trends Anal. Chem. 28 (10) (2009) 1201–1222. [23] A. Savitzky, M.J.E. Golay, Smoothing and differentiation of data by simplified least squares procedures, Anal. Chem. 36 (8) (1964) 1627–1639. [24] P.C. Williams, Implementation of near-infrared technology, in: P.C. Williams, K. Norris (Eds.), Near Infrared Technology in the Agriculture and Food Industries, 2nd ed, American Association of Cereal Chemists, Minnesota, USA, 2001. [25] C.C. Fagan, et al., Evaluating Mid-infrared spectroscopy as a new technique for predicting sensory texture attributes of processed cheese, J. Dairy Sci. 90 (3) (2007) 1122–1132. [26] C.M. Andersen, R. Bro, Variable selection in regression—a tutorial, J. Chemom. 24 (11 − 12) (2010) 728–737. [27] I.-G. Chong, C.-H. Jun, Performance of some variable selection methods when multicollinearity is present, Chemom. Intell. Lab. Syst. 78 (1–2) (2005) 103–112. [28] T. Rajalahti, et al., Biomarker discovery in mass spectral profiles by means of selectivity ratio plot, Chemom. Intell. Lab. Syst. 95 (1) (2009) 35–48. [29] L. Norgaard, et al., Interval Partial Least-Squares Regression (iPLS): a comparative chemometric study with an example from near-infrared spectroscopy, Appl. Spectrosc. 54 (3) (2000) 413–419. [30] Å. Rinnan, et al., Recursive weighted partial least squares (rPLS): an efficient variable selection method using PLS, J. Chemom. 28 (5) (2014) 439–447. [31] M. Schwanninger, J. Rodrigues, K. Fackler, A review of band assignments in near infrared spectra of wood and wood components, J. Near Infrared Spectrosc. 19 (5) (2011) 287–308. [32] C.-M. Popescu, M.-C. Popescu, A near infrared spectroscopic study of the structural modifications of lime (Tilia cordata Mill.) wood during hydro-thermal treatment, Spectrochim. Acta A Mol. Biomol. Spectrosc. 115 (2013) 227–233. [33] B.H. Stuart, Inorganic molecules, Infrared Spectroscopy: Fundamentals and Applications, 95-111 John Wiley & Sons, Ltd., 2005.
Acknowledgements The authors would like to thank Thermo Fisher Scientific, for the support in the development of this study. References [1] A. González, et al., Review of micro- and small-scale technologies to produce electricity and heat from Mediterranean forests' wood chips, Renew. Sust. Energ. Rev. 43 (0) (2015) 143–155. [2] M. Manzone, Quality, productivity, energy and costs of woodchip produced by Cedrus deodara plantations: a case study in Italy, Biomass Bioenergy 92 (2016) 81–87. [3] D.Y. Guimier, M.A. Pottie, Preparation of Forest Biomass for Optimal Conversion, Forest Engineering Research Institute of Canada, 1985. [4] G. Toscano, et al., Investigation of woodchip quality: relationship between the most important chemical and physical parameters, Energy 106 (2016) 38–44. [5] P. Dare, et al., Combustion performance of biomass residue and purpose grown species, Biomass Bioenergy 21 (4) (2001) 277–287. [6] B.R. Hames, et al., Rapid biomass analysis: New tools for compositional analysis of corn stover feedstocks and process intermediates from ethanol production, Appl. Biochem. Biotechnol. Part A Enzyme Eng. Biotechnol. 105 (1–3) (2003) 5–16. [7] S.V. Vassilev, et al., An overview of the chemical composition of biomass, Fuel 89 (5) (2010) 913–933. [8] 17225-4, E.I., solid biofuels - fuel specifications and classes - part 4, Graded Wood Chips, 2014. [9] J. Posom, et al., Rapid non-destructive evaluation of moisture content and higher heating value of Leucaena leucocephala pellets using near infrared spectroscopy, Energy 107 (2016) 464–472. [10] I. Lewandowski, A. Kicherer, Combustion quality of biomass: practical relevance and experiments to modify the biomass quality of Miscanthus × giganteus, Eur. J. Agron. 6 (3–4) (1997) 163–177. [11] M.K. Misra, K.W. Ragland, A.J. Baker, Wood ash composition as a function of furnace temperature, Biomass Bioenergy 4 (2) (1993) 103–116. [12] C.C. Fagan, C.D. Everard, K. McDonnell, Prediction of moisture, calorific value, ash and carbon content of two dedicated bioenergy crops using near-infrared spectroscopy, Bioresour. Technol. 102 (8) (2011) 5200–5206. [13] Gillespie, G.D., C.D. Everard, and K.P. McDonnell, Prediction of biomass pellet quality indices using near infrared spectroscopy. Energy, (0). [14] T.A. Lestander, C. Rhen, Multivariate NIR spectroscopy models for moisture, ash and calorific content in biofuels using bi-orthogonal partial least squares regression, Analyst 130 (8) (2005) 1182–1189. [15] C.D. Everard, K.P. McDonnell, C.C. Fagan, Prediction of biomass gross calorific values using visible and near infrared spectroscopy, Biomass Bioenergy 45 (0)
83