- Email: [email protected]
Prediction of biogas yield and its kinetics in reed canary grass using near infrared reflectance spectroscopy and chemometrics
Bioresource Technology 146 (2013) 282–287
Contents lists available at ScienceDirect
Bioresource Technology journal homepage: www.elsevier.com/locate/biortech
Prediction of biogas yield and its kinetics in reed canary grass using near infrared reflectance spectroscopy and chemometrics Tanka P. Kandel a,⇑, René Gislum b, Uffe Jørgensen a, Poul E. Lærke a a b
Department of Agroecology, Aarhus University Foulum, DK-8830 Tjele, Denmark Department of Agroecology, Aarhus University Flakkebjerg, DK-4200 Slagelse, Denmark
h i g h l i g h t s NIRS based models were developed for biogas yield, its kinetics, and methane yield. The iPLS models did not improve the models based on full spectrum data. The NIRS models for SBY and k-SBY were better than the model for SMY. NIRS model prediction for SMY was better than models based on chemical composition.
a r t i c l e
i n f o
Article history: Received 25 April 2013 Received in revised form 15 July 2013 Accepted 20 July 2013 Available online 26 July 2013 Keywords: Biogas iPLS PLS Methane NIR
a b s t r a c t A rapid method is needed to assess biogas and methane yield potential of various kinds of substrate prior to anaerobic digestion. This study reports near infrared reflectance spectroscopy (NIRS) as a rapid alternative method to the conventional batch methods for prediction of specific biogas yield (SBY), specific methane yield (SMY) and kinetics of biogas yield (k-SBY) of reed canary grass (RCG) biomass. Dried and powdered RCG biomass with different level of maturity was used for biochemical composition analysis, batch assays and NIRS analysis. Calibration models were developed using partial least square (PLS) regression from NIRS spectra. The calibration models for SBY (R2 = 0.68, RPD = 1.83) and k-SBY (R2 = 0.71, RPD = 1.75) were better than the model for SMY (R2 = 0.53, RPD = 1.49). Although the PLS model for SMY was less successful, the model performance was better compared to the models based on chemical composition. Ó 2013 Elsevier Ltd. All rights reserved.
1. Introduction Biogas production by anaerobic digestions of various agricultural crops and their residues is increasing worldwide as an alternative source of energy. Maize is the most common crop currently used for biogas production in European countries (Bruni et al., 2010; Grieder et al., 2011; Herrmann and Rath, 2012). However, perennial grasses are identified as a better option than maize as they generally have better environmental profile and require less input (Tilman et al., 2006). Moreover, cultivation of perennial grass also offers diversity in areas where monoculture of maize for biogas production is common. Reed canary grass (RCG) is identified as a suitable perennial crop to cultivate on peatlands in Denmark for biogas production (Kandel et al., 2012, 2013a,b). Biogas yields (volume of biogas produced per unit dry matter) of RCG vary considerably. It depends on chemical compositions of the biomass such as nitrogen concentrations, the amount of ⇑ Corresponding author. Tel.: +45 8715 4764; fax: +45 2343 1839. E-mail address: [email protected] (T.P. Kandel). 0960-8524/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.biortech.2013.07.092
structural and non-structural carbohydrates and lignification (Gunaseelan, 2007; Kandel et al., 2013a). The chemical composition is greatly influenced by factors such as variety, cultivation site, and agronomical managements including fertilization, harvest time and harvest frequency which eventually affect the biogas and methane yield (Seppälä et al., 2009; Kandel et al., 2013a). Therefore, it is important to determine the biogas and methane yield potentials of RCG prior to anaerobic digestion in order to assess its value as feedstock and also to obtain information on required feeding rate and retention time in the biogas reactor (Raju et al., 2011). Assessment of the biochemical methane potential (BMP) by a lab procedure with small anaerobic reactors is the most common method to assess the biogas yield potential of various feedstocks (Møller et al., 2004; Grieder et al., 2011). However, it takes at least 30 days to assess the anaerobic digestion potential by this batch method. Biomass with higher percentage of fibrous components can even take longer time (Triolo et al., 2011; Kandel et al., 2013a). Moreover, the test is also labor and resource intensive (Grieder et al., 2011). As the BMP assay is time consuming, it is
283
T.P. Kandel et al. / Bioresource Technology 146 (2013) 282–287
not a practical method for assessing biogas yield potential of biomass at industrial scale (Lesteur et al., 2011). Therefore, a more rapid method is needed to assess the quantity of biogas and methane that may be produced from different kinds of substrate as the biogas industry is booming in Northern Europe and other parts of the world. Various methods have been used to assess the biogas production as an alternative to the BMP assays (Lesteur et al., 2010). Modelling biogas or methane yield based on chemical composition of the biomass is one of the common methods (Gunaseelan, 2007; Triolo et al., 2011). The BMP models based on chemical composition analysis can estimate the biogas or methane yield in shorter duration as compared to the fermentation assays and they provide an insight on factors responsible for decomposition of substrates. However, modelling still needs information on the chemical composition itself, and this analysis can be expensive and labor intensive. In vitro organic matter digestibility assay (IVOMD) (Raju et al., 2011) and Dynamic Respiration Index (DRI) (Scaglia et al., 2010) are relatively faster methods as compared to BMP assays but they are also time and labor intensive methods. Near infrared reflectance spectroscopy (NIRS) may be a method for faster assessment of anaerobic digestion potential of feedstocks (Raju et al., 2011; Lesteur et al., 2011; Grieder et al., 2011). NIRS is a non-invasive method, and time and labor requirement in this method is significantly lower as compared to the BMP assays. NIRS has already been proved to be a powerful tool to determine the chemical composition of various plants parts (Gislum et al., 2004; Shetty and Gislum, 2011). As the chemical composition of the biomass determines BMP of the feedstocks and NIRS can determine the chemical composition, some attempts were already made to directly predict BMP from various kinds of biomass using NIRS. Those studies include BMP potential of municipal solid wastes (Lesteur et al., 2011), meadow grass (Raju et al., 2011) and maize (Grieder et al., 2011). Although these studies have shown certain ability of NIRS on predicting BMP, further studies are needed to make the technique applicable to a wider ranges of biomasses and to identify the reasons to lack of good model fitting. This study investigated NIRS as a tool to predict specific biogas yield (SBY), specific methane yield (SMY), and kinetics of biogas yield (k-SBY) of RCG biomass with different level of maturity and chemical composition.
2. Methods 2.1. Plant materials The plant materials used for this study were 98 samples of RCG from an earlier study (Kandel et al., 2013a). RCG was grown on a peatland located in Nørre Å river valley near to Viborg, Denmark. Environmental conditions during crop cultivation, soil properties and design of the field experiment are described in details by Kandel et al. (2012, 2013b). The experimental site consisted of three plots of RCG established in 2009 which were divided into two subplots (18 12 m). Biomass was sampled (1 m2 area in each sampling) from those subplots in 2 week intervals from April to September in 2011. Similarly, regrowth of biomass in one of the summer harvests was harvested again in September for a twocut management strategy with or without additional fertilization after the first cut. The biomass was separated into leaf and stem, and oven dried to constant weight. Subsequently, the biomass was ground in a mill to pass through a 1 mm sieve. The ground samples were divided into three parts used for biochemical composition analysis, BMP assays and NIRS analysis.
2.2. Biogas reference data Biochemical methane potential (BMP) of RCG was determined using a method described by Møller et al. (2004). In brief, 4 g of dry, ground sample was digested at mesophillic condition (35 °C) in a batch reactor (500 ml) filled with inoculum (200 g) obtained from a mesophillic (25 °C) post digestion tank. Three batch reactors without addition of RCG were used as control. The volume of the biogas produced was measured every week at the beginning of the batch assay and then gradually at longer time intervals using an acidified (pH < 2) water displacement method. Total biogas produced from the biomass samples during the batch assay period (69 days) were corrected to normal conditions (0 °C and 1.013 bar) and presented as specific biogas yield (SBY) [NL biogas (kg VS)1]. Methane concentration in the biogas was analyzed by gas chromatography (HP 6890 series; Agilent Technologies, Copenhagen, Denmark). Total methane yield was calculated as the product of SBY and methane concentration and presented as specific methane yield (SMY) [NL CH4 (kg VS)1]. To determine the kinetics of biogas yield, a simple first order degradation model (e.g., Gunaseelan, 2007) was used for measurements from individual batch reactors:
B ¼ B0 ð1 expkt Þ
ð1Þ
where B is the cumulative biogas yield at time t, B0 is the ultimate biogas yield and k is the first order rate constant. Here, k is used as kinetics of biodegradation or biogas yield and presented as k-SBY. Biomass with low k-SBY degrade over a long time period with lower rate of biogas production while the biomass with high k-SBY degrade immediately in batch reactor with high rates of biogas formation at the beginning of the batch assay. 2.3. Near infrared reflectance spectroscopy The grass samples were measured by NIRS spectroscopy (Q-interline Spectroscopic Analytical Solutions, QFAflex, Roskilde, Denmark) in reflectance mode where log (1/R) was recorded from 1100 to 2498 nm, equivalent to 4000–9091 cm1. Each sample was scanned 128 times using 16 cm1 resolution. 2.4. Multivariate data analysis Exploratory data analysis of all samples using principal component analysis (PCA) and leverage plots of the NIRS spectra revealed a few obvious outliers which were removed before partial least squares regression (PLS) models were developed on full spectra. The full dataset was divided into a calibration and validation dataset where data were arranged in increasing order and every third sample was used for test set validation. Data in the test set was
Table 1 Reference data parameters. Specific methane yield (SMY), specific biogas yield (SBY) and kinetics of biogas yield (k-SBY). Parameters
SBY
SMY
k-SBY
Calibration n Min Max Mean S. Dev
69 431 699 564 68
69 264 401 337 32
69 0.038 0.082 0.059 0.010
Validation n Min Max Mean S. Dev
22 440 673 563 55
22 272 396 338 29
22 0.040 0.087 0.061 0.011
284
T.P. Kandel et al. / Bioresource Technology 146 (2013) 282–287
Different pre-processing methods were applied to the raw spectra to correct for scatter; 1st derivative (1d), 2nd derivative (2d), multiplicative scatter correction (MSC) (Geladi et al., 1985) and standard normal variate (SNV) (Barnes et al., 1989). 1d and 2d NIRS spectra were obtained using Savitzky Golay (Savitzky and Golay, 1964) with 2nd order polynomial and smoothing over 7 units. For each PLS model the optimum pre-processing method of the NIRS data and the optimum number of latent variables (LV) was chosen using root mean square error of cross validation (RMSECV) plotted against number of LV, where the optimum number was chosen when the curve started to flatten. The PLS models were validated using random segmented cross validation with 10 data splits and 9 interactions to overcome effect of random selection. The other method used for modelling was forward interval PLS (iPLS) models (Nørgaard et al., 2000). The iPLS models were developed using 10 or 20 equidistant sub-intervals. The iPLS algorithm will continue to add or remove intervals until the cross-validation results cease to improve. The iPLS models were based on the best pre-processing spectra of the PLS models. Performances of the models were evaluated by correlation coefficients (R2) and root mean square error (RMSE), bias, and residual prediction deviation (RPD). The R2 defines the proportion of variation explained by the model, the RMSE indicated the average error of prediction, bias estimate the systematic error, and the RPD is the ratio of standard deviation to RMSE. A model with higher R2 and RPD and lower RMSE is considered a successful model. All calculations were made using Matlab version 7.8.0.347, 2009a (MathWorks Inc., Natick, MA, USA) and PLS toolbox version 7.0.1, 2012 (Eigenvector Research Inc., Wenatchee, WA, USA).
RMSECV [NL biogas (kg VS)-1]
100
Specific Biogas yield (SBY) Raw 1d 2d MSC 1dMSC
80
60
40
RMSECV [NL CH4 (kg VS)-1]
Specific methane yield (SMY) 50
40
30
20
RMSECV (day-1)
0.016
Biogas production rate constant (k-SBY)
0.012
3. Results and discussion
0.008
3.1. Choice of the best pre-processing method 0.004 0
5
10
15
20
Latent variables (LV) Fig. 1. RMSECV vs. optimum number of latent variables using random subset cross validation with different pre-processing methods.
therefore not used in the development of the calibration model. The reference data used for calibration and validation are tabulated in Table 1.
All scatter correction methods lowered the RMSECV compared to raw data at 3 latent variables (LV) for both SMY and SBY and 5 LV for k-SBY (Fig. 1). In the higher CV levels, RMSECV increased in all methods. A sharp decrease in RMSECV was observed in 1d and 2d pre-processing methods which indicated that variation present in first CV is well focused on Y variables. The preprocessing methods chosen for SBY, SMY and k-SBY and the respective number of latent variables are presented in Table 2. For further modelling, 2d method was chosen for SBY and k-SBY models and 1d method was used for SMY.
Table 2 Calibration and validation statistics of PLS and iPLS models with different pre-processing methods. Best PLS models are marked as bold. For iPLS, models were based on the best preprocessing methods of PLS best models. Calibration model
Validation model
Parameter
Pre-processing
No. of LV
R2
RMSECV
RPDSDc/RMSECV
R2
RMSEP
Bias
PLS model SBY SBY SBY SBY SMY SMY k-SBY k-SBY k-SBY k-SBY
1d 2d 2d 1d 1d 2d 1d 2d 2d 2d
4 3 4 3 3 3 3 3 5 7
0.69 0.65 0.70 0.68 0.52 0.53 0.66 0.64 0.71 0.76
36.7 39.0 36.2 37.3 22.4 21.6 0.0062 0.0063 0.0057 0.0052
1.86 1.75 1.89 1.83 1.44 1.49 1.61 1.58 1.75 1.92
0.56 0.41 0.54 0.60 0.26 0.26 0.64 0.67 0.87 0.86
41.4 48.3 42.7 40.0 25.3 24.7 0.0057 0.0055 0.0034 0.0035
3.43 0.49 2.28 3.75 0.61 0.004 0.0011 0.0013 0.0001 0.0008
iPLS model SBY SMY k-SBY
1d 2d 2d
3 3 5
0.64 0.52 0.71
39.3 22.2 0.0057
1.83 1.49 1.75
0.53 0.11 0.84
42.7 30.0 0.0038
2.47 1.68 0.0001
285
T.P. Kandel et al. / Bioresource Technology 146 (2013) 282–287
3.2. Comparison of PLS model with previous studies Previous studies (Grieder et al., 2011; Lesteur et al., 2011; Raju et al., 2011) have reported the NIRS based prediction of SMY but they did not report SBY as those studies did not show large difference in CH4 concentrations among samples. The PLS model used to predict SMY in this study was less successful as compared to these previous studies (Table 3). All three previous studies have reported R2 values of calibration and validation models in the range of 0.69– 0.79 which are considerably higher as compared to this study. Similarly, RPD values of NIRS model in this study is slightly lower than the previous studies. Raju et al. (2011) suggested using homogenous samples in batch assays to improve SMY models based on NIRS spectroscopy. Samples used for the batch assay and the NIRS spectra measurements were powdered and homogenized in this
study as suggested but the model performance was not improved. Raju et al. (2011) also suggested using a wider range of SMY values in the reference data set for PLS model calibration to improve SMY prediction. The reference data used in this study had a wide range of chemical composition (Kandel et al., 2013a) as the biomass was harvested regularly from the beginning of the growing season to maturity. As expected, the biomass with lower concentration of structural components produced higher amount of biogas but lower methane concentrations in the biogas produced from the biomass offset the total methane yield. This resulted in a larger range of SBY compared to SMY in the reference data (Table 1). The reference data used in this study had a relatively smaller range of SMY as compared to Lesteur et al. (2011) and Raju et al. (2011) but the range was wider than the data used by Grieder et al. (2011)
0.3
Specific Biogas Yield (SBY) LV1 (98.33) LV2 (0.89) LV3 (0.44)
Loading Weights
0.2
0.1
0.0
-0.1
0.3
Specific Methane Yield (SMY) 0.2 LV1 (96.18%) LV2 (3.11%) LV3 (0.39%)
Loading Weights
0.1
0.0
-0.1
-0.2
-0.3
-0.4
Kinetics of Specific Biogas Yield (k-SBY) 0.2
0.1
Loading Weights
The iPLS models did not improve the models based on full spectrum data for any of the studied parameters (SBY, SMY and k-SBY) (Table 2). When there is high degree of redundancy in NIRS spectra, the iPLS method can improve full-spectrum PLS models as iPLS method eliminates the redundancy (Nørgaard et al., 2000). Therefore, the lower prediction ability of iPLS models in this study probably indicates lower redundancy in the NIRS spectra, meaning that the PLS models use information from large part of the NIRS spectra. As the iPLS models did not improve the model performance, the PLS models based on full spectrum data were used for further comparisons. Loading plots for latent variables are shown in Fig. 2. The loading plot makes it possible to identify the spectral regions that are important for the models. The loading plot (Fig. 2) shows a very complex pattern and that various spectral regions from 1800 to 2500 nm are important. This is also supported by lower prediction ability of the iPLS model compared to the model based on full spectrum data. However, the first latent variable (LV1) which explains >95% variations in Y-data have the highest absolute value around 1800–1900 nm spectral region. Previous studies have shown that the spectral region around 1900 nm describe the lignin content of the biomass (Uner et al., 2009). Similarly, spectral region around 1900 and 1930 are one of the few regions assigned to cellulose (Jones et al., 2006). Previous studies have also shown that chemical composition such as lignin and cellulose concentration of the biomass can predict SBY and SMY (Grieder et al., 2011; Triolo et al., 2011; Kandel et al., 2013a). As the spectral regions, that are important for lignin and cellulose concentration, also explain SBY and SMY to some extent, it is clear that NIRS could directly predict biodegradability of various kinds of biomass. Performance indicator statistics of the PLS models with the best pre-processing methods are also presented in Table 2. The corresponding measured versus predicted plots are presented in Fig. 3. The SBY and k-SBY models were better than the SMY for both the calibration and validation models. The better prediction of SBY by PLS models suggests that the information present in NIRS spectroscopy data explained biodegradability of biomass better but the composition of the final products (concentration of CO2 and CH4) of the digestion was not be explained well. Grieder et al. (2011) also reported that SBY was predicted successfully by NIRS model but the model was less successful in predicting methane concentration dynamics during the batch assay. The reference data used in this study had large difference in proportion of CH4 and CO2 in the biogas which may be difficult to explain by information obtained from the NIRS spectroscopy data. Also, models based on chemical composition of the biomass had predicted SBY better as compared to SMY (Kandel et al., 2013a). The better performance of k-SBY indicates that the rate of degradation of biomass could be predicted well with NIRS model which is important for determination of the feeding rate into the biogas reactor.
0.0
-0.1
LV1 (96.11%) LV2 (3.18%) LV3 (0.38%) LV4 (0.07%) LV5 (0.07%)
-0.2
-0.3
-0.4 1200
1400
1600
1800
2000
2200
2400
Variables Fig. 2. Loading plots for latent variables (LVs). Optimum numbers of LVs were three for SBY and SMY, and five for k-SBY. Variations represented by the latent variables (LVs) are given in percentages inside bracket.
T.P. Kandel et al. / Bioresource Technology 146 (2013) 282–287
680
620
560
500
440
420
0.09
390
0.08 -1
Modelled SMY [NL CH4 (kg VS)-1]
Modelled SBY [NL biogas (kg VS)-1]
740
Modelled k-SBY (day )
286
360 330 300 270
440
500
560
620
680
740
240
Measured SBY [NL biogas (kg VS)-1]
0.06 0.05 0.04
240
380 380
0.07
0.03 270
300
330
360
390
0.03
420
0.04
0.05
0.06
0.07
0.08
0.09
Measured k-SBY (day-1)
Measured SMY [NL CH4 (kg VS)-1]
Fig. 3. Scatter plots of measured vs. predicted for the PLS models. Closed circles represent data from calibration models and open circles represent data from validation models.
Table 3 Comparison of PLRS model performances of specific methane yield (SMY) in this study with previous studies. Study
Calibration model
Validation model
Performance statistics
Grieder et al. (2011) Lesteur et al. (2011) Raju et al. (2011)a This study a
SMY (Reference data)
Performance statistics
SMY (Reference data)
R2
RMSEcv
RPD
Mean
SD
Range
R2
RMSEP
RPD
Mean
SD
Range
0.76 0.79 0.69 0.53
6.2 31 37.4 21.6
2.07 2.13 1.75 1.49
329 234 288 337
13 66 66 32
295–355 23–400 51–406 264–401
0.77 0.76
6.4 28.0
2.08 2.36
329 227
13 57
301–357 87–322
0.26
24.7
1.20
338
29
272–396
Did not present different calibration and validation models.
Table 4 Comparison of different models of specific methane yield (SMY) based on chemical composition and PLS methods. Model
R2
RMSE
Bias
RPD
Lignin ? SMY Cellulose ? SMY NDF ? SMY ADF ? SMY Lignin + Cellulose ? SMY NIRS ? SMY
0.36 0.32 0.42 0.35 0.37 0.53
26.5 27.4 25.2 26.7 26.3 21.6
0.00 0.01 0.01 0.01 0.01 0.00
1.47 1.42 1.54 1.46 1.48 1.49
(Table 3). The reason for lower success of PLS model in this study as compared to the previous studies is not clear. However, as stated earlier, the difference in proportion of CH4 and CO2 in the biogas from the different samples might be one reason for poor performance as the SBY model shows performance similar to SMY models in previous studies (Table 1).
4. Conclusions This study reported NIRS based prediction of specific biogas yield, kinetics of biogas yield, and specific methane yield of RCG biomass. The iPLS models did not improve the models based on full spectrum data. Various spectral regions from 1800 to 2500 nm were important for prediction efficiency indicating the importance of fiber fractions such as cellulose and lignin concentrations. The calibration and validation models for SBY and k-SBY were better than the model for SMY. Although NIRS model presented in this study was less successful compared to the previous studies, the model prediction was better than the models based on chemical composition. Acknowledgements The study was supported by the European Regional Development Fund as a part of the projects ENERCOAST (http://enercoast.net) and BioM (http://www.biom-kask.eu).
3.3. Comparison of the PLS model with models based on chemical composition
References
Comparison of performance of SMY models based on chemical composition of RCG with performance of the PLS model is presented in Table 4. The PLS model showed better performance as compared to the models based on chemical components. A reason for this could be that the NIRS spectra may contain information not only about the composition of the biomass, such as lignin and cellulose content, but it also contains information about associations among such components (Grieder et al., 2011). Such additional information that can be present in NIRS spectra may have contributed to the better NIRS model as compared to the models based on chemical composition. However, the PLS model is still not strong enough to predict the methane yield with a satisfactory precision.
Barnes, R.J., Dhanoa, M.S., Lister, S.J., 1989. Standard normal variate transformation and de-trending of near-infrared diffuse reflectance spectra. Appl. Spectrosc. 43, 772–777. Bruni, E., Jensen, A.P., Pedersen, E.S., Angelidaki, I., 2010. Anaerobic digestion of maize focusing on variety, harvest time and pretreatment. Appl. Energy 87, 2212–2217. Geladi, P., McDougall, D., Martens, H., 1985. Linearization and scatter-correction for near-infrared reflectance spectra of meat. Appl. Spectrosc. 39, 491–500. Gislum, R., Micklander, E., Nielsen, J.P., 2004. Quantification of nitrogen concentration in perennial ryegrass and red fescue using near-infrared reflectance spectroscopy (NIRS) and chemometrics. Field Crops Res. 88, 269– 277. Grieder, C., Mittweg, G., Dhillon, B.S., Montes, J.M., Orsini, E., Melchinger, A.E., 2011. Determination of methane fermentation yield and its kinetics by near infrared spectroscopy and chemical composition in maize. J. Near Infrared Spectrosc. 19, 463–477.
T.P. Kandel et al. / Bioresource Technology 146 (2013) 282–287 Gunaseelan, V.N., 2007. Regression models of ultimate methane yields of fruits and vegetable solid wastes, sorghum and napiergrass on chemical composition. Bioresour. Technol. 98, 1270–1277. Herrmann, A., Rath, J., 2012. Biogas production from maize: current state, challenges, and prospects. 1. Methane yield potential. Bioenergy Res. 5, 1027– 1042. Jones, P.D., Schimleck, L.R., Peter, G.F., Daniels, R.F., Clark III, A., 2006. Nondestructive estimation of wood chemical composition of sections of radial wood strips by diffuse reflectance near infrared spectroscopy. Wood Sci. Technol. 40 (8), 709–720. Kandel, T.P., Elsgaard, L., Lærke, P.E., 2012. Measurement and modelling of CO2 flux from a drained fen peatland cultivated with reed canary grass and spring barley. Global Change Biol. Bioenergy. http://dx.doi.org/10.1111/gcbb.12020. Kandel, T.P., Sutaryo, S., Møller, H.B., Jørgensen, U., Lærke, P.E., 2013a. Chemical composition and methane yield of reed canary grass as influenced by harvesting time and harvest frequency. Bioresour. Technol. 130, 659–666. Kandel, T.P., Elsgaard, L., Karki, S., Lærke, P.E., 2013b. Biomass yield and greenhouse gas emissions from a drained fen peatland cultivated with reed canary grass under different harvest and fertilizer regimes. Bioenergy Res.. http://dx.doi.org/ 10.1007/s12155-013-9316-5. Lesteur, M., Bellon-Maurel, V., Gonzalez, C., Latrille, E., Roger, J.M., Junqua, G., Steyer, J.P., 2010. Alternative methods for determining anaerobic biodegradability: a review. Process Biochem. 45, 431–440. Lesteur, M., Latrille, E., Maurel, V.B., Roger, J.M., Gonzalez, C., Junqua, G., Steyer, J.P., 2011. First step towards a fast analytical method for the determination of biochemical methane potential of solid wastes by near infrared spectroscopy. Bioresour. Technol. 102, 2280–2288. Møller, H.B., Sommer, S.G., Ahring, B.K., 2004. Methane productivity of manure, straw and solid fractions of manure. Biomass Bioenergy 26, 485–495.
287
Nørgaard, L., Saudland, A., Wagner, J., Nielsen, J.P., Munck, L., Engelsen, S.B., 2000. Interval partial least-squares regression (iPLS): a comparative chemometric study with an example from near-infrared spectroscopy. Appl. Spectrosc. 54, 413–419. Raju, C.S., Ward, A.J., Nielsen, L., Møller, H.B., 2011. Comparison of near infra-red spectroscopy, neutral detergent fiber assay and in-vitro organic matter digestibility assay for rapid determination of the biochemical methane potential of meadow grasses. Bioresour. Technol. 102, 7835–7839. Savitzky, A., Golay, M.J.E., 1964. Smoothing and differentiation of data by simplified least squares procedures. Anal. Chem. 36, 1627–1638. Scaglia, B., Confalonieri, R., D’Imporzano, G., Adani, F., 2010. Estimating biogas production of biologically treated municipal solid waste. Bioresour. Technol. 101, 945–952. Seppälä, M., Paavola, T., Lehtomäki, A., Rintala, J., 2009. Biogas production from boreal herbaceous grasses – specific methane yield and methane yield per hectare. Bioresour. Technol. 100, 2952–2958. Shetty, N., Gislum, R., 2011. Quantification of fructan concentration in grasses using NIR spectroscopy and PLSR. Field Crops Res. 120, 31–37. Tilman, D., Hill, J., Lehman, C., 2006. Carbon-negative biofuels from low-input highdiversity grassland biomass. Science 314, 1598–1600. Triolo, J.M., Sommer, S.G., Møller, H.B., Weisbjerg, M.R., Jiang, X.Y., 2011. A new algorithm to characterize biodegradability of biomass during anaerobic digestion: Influence of lignin concentration on methane production potential. Bioresour. Technol. 102, 9395–9402. Uner, B., Karaman, I., Tanriverdi, H., Özdemir, D., 2009. Prediction of lignin and extractive content of Pinus nigra Arnold. var. Pallasiana tree using near infrared spectroscopy and multivariate calibration. J. Wood Chem. Technol. 29 (1), 24–42.
Contents lists available at ScienceDirect
Bioresource Technology journal homepage: www.elsevier.com/locate/biortech
Prediction of biogas yield and its kinetics in reed canary grass using near infrared reflectance spectroscopy and chemometrics Tanka P. Kandel a,⇑, René Gislum b, Uffe Jørgensen a, Poul E. Lærke a a b
Department of Agroecology, Aarhus University Foulum, DK-8830 Tjele, Denmark Department of Agroecology, Aarhus University Flakkebjerg, DK-4200 Slagelse, Denmark
h i g h l i g h t s NIRS based models were developed for biogas yield, its kinetics, and methane yield. The iPLS models did not improve the models based on full spectrum data. The NIRS models for SBY and k-SBY were better than the model for SMY. NIRS model prediction for SMY was better than models based on chemical composition.
a r t i c l e
i n f o
Article history: Received 25 April 2013 Received in revised form 15 July 2013 Accepted 20 July 2013 Available online 26 July 2013 Keywords: Biogas iPLS PLS Methane NIR
a b s t r a c t A rapid method is needed to assess biogas and methane yield potential of various kinds of substrate prior to anaerobic digestion. This study reports near infrared reflectance spectroscopy (NIRS) as a rapid alternative method to the conventional batch methods for prediction of specific biogas yield (SBY), specific methane yield (SMY) and kinetics of biogas yield (k-SBY) of reed canary grass (RCG) biomass. Dried and powdered RCG biomass with different level of maturity was used for biochemical composition analysis, batch assays and NIRS analysis. Calibration models were developed using partial least square (PLS) regression from NIRS spectra. The calibration models for SBY (R2 = 0.68, RPD = 1.83) and k-SBY (R2 = 0.71, RPD = 1.75) were better than the model for SMY (R2 = 0.53, RPD = 1.49). Although the PLS model for SMY was less successful, the model performance was better compared to the models based on chemical composition. Ó 2013 Elsevier Ltd. All rights reserved.
1. Introduction Biogas production by anaerobic digestions of various agricultural crops and their residues is increasing worldwide as an alternative source of energy. Maize is the most common crop currently used for biogas production in European countries (Bruni et al., 2010; Grieder et al., 2011; Herrmann and Rath, 2012). However, perennial grasses are identified as a better option than maize as they generally have better environmental profile and require less input (Tilman et al., 2006). Moreover, cultivation of perennial grass also offers diversity in areas where monoculture of maize for biogas production is common. Reed canary grass (RCG) is identified as a suitable perennial crop to cultivate on peatlands in Denmark for biogas production (Kandel et al., 2012, 2013a,b). Biogas yields (volume of biogas produced per unit dry matter) of RCG vary considerably. It depends on chemical compositions of the biomass such as nitrogen concentrations, the amount of ⇑ Corresponding author. Tel.: +45 8715 4764; fax: +45 2343 1839. E-mail address: [email protected] (T.P. Kandel). 0960-8524/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.biortech.2013.07.092
structural and non-structural carbohydrates and lignification (Gunaseelan, 2007; Kandel et al., 2013a). The chemical composition is greatly influenced by factors such as variety, cultivation site, and agronomical managements including fertilization, harvest time and harvest frequency which eventually affect the biogas and methane yield (Seppälä et al., 2009; Kandel et al., 2013a). Therefore, it is important to determine the biogas and methane yield potentials of RCG prior to anaerobic digestion in order to assess its value as feedstock and also to obtain information on required feeding rate and retention time in the biogas reactor (Raju et al., 2011). Assessment of the biochemical methane potential (BMP) by a lab procedure with small anaerobic reactors is the most common method to assess the biogas yield potential of various feedstocks (Møller et al., 2004; Grieder et al., 2011). However, it takes at least 30 days to assess the anaerobic digestion potential by this batch method. Biomass with higher percentage of fibrous components can even take longer time (Triolo et al., 2011; Kandel et al., 2013a). Moreover, the test is also labor and resource intensive (Grieder et al., 2011). As the BMP assay is time consuming, it is
283
T.P. Kandel et al. / Bioresource Technology 146 (2013) 282–287
not a practical method for assessing biogas yield potential of biomass at industrial scale (Lesteur et al., 2011). Therefore, a more rapid method is needed to assess the quantity of biogas and methane that may be produced from different kinds of substrate as the biogas industry is booming in Northern Europe and other parts of the world. Various methods have been used to assess the biogas production as an alternative to the BMP assays (Lesteur et al., 2010). Modelling biogas or methane yield based on chemical composition of the biomass is one of the common methods (Gunaseelan, 2007; Triolo et al., 2011). The BMP models based on chemical composition analysis can estimate the biogas or methane yield in shorter duration as compared to the fermentation assays and they provide an insight on factors responsible for decomposition of substrates. However, modelling still needs information on the chemical composition itself, and this analysis can be expensive and labor intensive. In vitro organic matter digestibility assay (IVOMD) (Raju et al., 2011) and Dynamic Respiration Index (DRI) (Scaglia et al., 2010) are relatively faster methods as compared to BMP assays but they are also time and labor intensive methods. Near infrared reflectance spectroscopy (NIRS) may be a method for faster assessment of anaerobic digestion potential of feedstocks (Raju et al., 2011; Lesteur et al., 2011; Grieder et al., 2011). NIRS is a non-invasive method, and time and labor requirement in this method is significantly lower as compared to the BMP assays. NIRS has already been proved to be a powerful tool to determine the chemical composition of various plants parts (Gislum et al., 2004; Shetty and Gislum, 2011). As the chemical composition of the biomass determines BMP of the feedstocks and NIRS can determine the chemical composition, some attempts were already made to directly predict BMP from various kinds of biomass using NIRS. Those studies include BMP potential of municipal solid wastes (Lesteur et al., 2011), meadow grass (Raju et al., 2011) and maize (Grieder et al., 2011). Although these studies have shown certain ability of NIRS on predicting BMP, further studies are needed to make the technique applicable to a wider ranges of biomasses and to identify the reasons to lack of good model fitting. This study investigated NIRS as a tool to predict specific biogas yield (SBY), specific methane yield (SMY), and kinetics of biogas yield (k-SBY) of RCG biomass with different level of maturity and chemical composition.
2. Methods 2.1. Plant materials The plant materials used for this study were 98 samples of RCG from an earlier study (Kandel et al., 2013a). RCG was grown on a peatland located in Nørre Å river valley near to Viborg, Denmark. Environmental conditions during crop cultivation, soil properties and design of the field experiment are described in details by Kandel et al. (2012, 2013b). The experimental site consisted of three plots of RCG established in 2009 which were divided into two subplots (18 12 m). Biomass was sampled (1 m2 area in each sampling) from those subplots in 2 week intervals from April to September in 2011. Similarly, regrowth of biomass in one of the summer harvests was harvested again in September for a twocut management strategy with or without additional fertilization after the first cut. The biomass was separated into leaf and stem, and oven dried to constant weight. Subsequently, the biomass was ground in a mill to pass through a 1 mm sieve. The ground samples were divided into three parts used for biochemical composition analysis, BMP assays and NIRS analysis.
2.2. Biogas reference data Biochemical methane potential (BMP) of RCG was determined using a method described by Møller et al. (2004). In brief, 4 g of dry, ground sample was digested at mesophillic condition (35 °C) in a batch reactor (500 ml) filled with inoculum (200 g) obtained from a mesophillic (25 °C) post digestion tank. Three batch reactors without addition of RCG were used as control. The volume of the biogas produced was measured every week at the beginning of the batch assay and then gradually at longer time intervals using an acidified (pH < 2) water displacement method. Total biogas produced from the biomass samples during the batch assay period (69 days) were corrected to normal conditions (0 °C and 1.013 bar) and presented as specific biogas yield (SBY) [NL biogas (kg VS)1]. Methane concentration in the biogas was analyzed by gas chromatography (HP 6890 series; Agilent Technologies, Copenhagen, Denmark). Total methane yield was calculated as the product of SBY and methane concentration and presented as specific methane yield (SMY) [NL CH4 (kg VS)1]. To determine the kinetics of biogas yield, a simple first order degradation model (e.g., Gunaseelan, 2007) was used for measurements from individual batch reactors:
B ¼ B0 ð1 expkt Þ
ð1Þ
where B is the cumulative biogas yield at time t, B0 is the ultimate biogas yield and k is the first order rate constant. Here, k is used as kinetics of biodegradation or biogas yield and presented as k-SBY. Biomass with low k-SBY degrade over a long time period with lower rate of biogas production while the biomass with high k-SBY degrade immediately in batch reactor with high rates of biogas formation at the beginning of the batch assay. 2.3. Near infrared reflectance spectroscopy The grass samples were measured by NIRS spectroscopy (Q-interline Spectroscopic Analytical Solutions, QFAflex, Roskilde, Denmark) in reflectance mode where log (1/R) was recorded from 1100 to 2498 nm, equivalent to 4000–9091 cm1. Each sample was scanned 128 times using 16 cm1 resolution. 2.4. Multivariate data analysis Exploratory data analysis of all samples using principal component analysis (PCA) and leverage plots of the NIRS spectra revealed a few obvious outliers which were removed before partial least squares regression (PLS) models were developed on full spectra. The full dataset was divided into a calibration and validation dataset where data were arranged in increasing order and every third sample was used for test set validation. Data in the test set was
Table 1 Reference data parameters. Specific methane yield (SMY), specific biogas yield (SBY) and kinetics of biogas yield (k-SBY). Parameters
SBY
SMY
k-SBY
Calibration n Min Max Mean S. Dev
69 431 699 564 68
69 264 401 337 32
69 0.038 0.082 0.059 0.010
Validation n Min Max Mean S. Dev
22 440 673 563 55
22 272 396 338 29
22 0.040 0.087 0.061 0.011
284
T.P. Kandel et al. / Bioresource Technology 146 (2013) 282–287
Different pre-processing methods were applied to the raw spectra to correct for scatter; 1st derivative (1d), 2nd derivative (2d), multiplicative scatter correction (MSC) (Geladi et al., 1985) and standard normal variate (SNV) (Barnes et al., 1989). 1d and 2d NIRS spectra were obtained using Savitzky Golay (Savitzky and Golay, 1964) with 2nd order polynomial and smoothing over 7 units. For each PLS model the optimum pre-processing method of the NIRS data and the optimum number of latent variables (LV) was chosen using root mean square error of cross validation (RMSECV) plotted against number of LV, where the optimum number was chosen when the curve started to flatten. The PLS models were validated using random segmented cross validation with 10 data splits and 9 interactions to overcome effect of random selection. The other method used for modelling was forward interval PLS (iPLS) models (Nørgaard et al., 2000). The iPLS models were developed using 10 or 20 equidistant sub-intervals. The iPLS algorithm will continue to add or remove intervals until the cross-validation results cease to improve. The iPLS models were based on the best pre-processing spectra of the PLS models. Performances of the models were evaluated by correlation coefficients (R2) and root mean square error (RMSE), bias, and residual prediction deviation (RPD). The R2 defines the proportion of variation explained by the model, the RMSE indicated the average error of prediction, bias estimate the systematic error, and the RPD is the ratio of standard deviation to RMSE. A model with higher R2 and RPD and lower RMSE is considered a successful model. All calculations were made using Matlab version 7.8.0.347, 2009a (MathWorks Inc., Natick, MA, USA) and PLS toolbox version 7.0.1, 2012 (Eigenvector Research Inc., Wenatchee, WA, USA).
RMSECV [NL biogas (kg VS)-1]
100
Specific Biogas yield (SBY) Raw 1d 2d MSC 1dMSC
80
60
40
RMSECV [NL CH4 (kg VS)-1]
Specific methane yield (SMY) 50
40
30
20
RMSECV (day-1)
0.016
Biogas production rate constant (k-SBY)
0.012
3. Results and discussion
0.008
3.1. Choice of the best pre-processing method 0.004 0
5
10
15
20
Latent variables (LV) Fig. 1. RMSECV vs. optimum number of latent variables using random subset cross validation with different pre-processing methods.
therefore not used in the development of the calibration model. The reference data used for calibration and validation are tabulated in Table 1.
All scatter correction methods lowered the RMSECV compared to raw data at 3 latent variables (LV) for both SMY and SBY and 5 LV for k-SBY (Fig. 1). In the higher CV levels, RMSECV increased in all methods. A sharp decrease in RMSECV was observed in 1d and 2d pre-processing methods which indicated that variation present in first CV is well focused on Y variables. The preprocessing methods chosen for SBY, SMY and k-SBY and the respective number of latent variables are presented in Table 2. For further modelling, 2d method was chosen for SBY and k-SBY models and 1d method was used for SMY.
Table 2 Calibration and validation statistics of PLS and iPLS models with different pre-processing methods. Best PLS models are marked as bold. For iPLS, models were based on the best preprocessing methods of PLS best models. Calibration model
Validation model
Parameter
Pre-processing
No. of LV
R2
RMSECV
RPDSDc/RMSECV
R2
RMSEP
Bias
PLS model SBY SBY SBY SBY SMY SMY k-SBY k-SBY k-SBY k-SBY
1d 2d 2d 1d 1d 2d 1d 2d 2d 2d
4 3 4 3 3 3 3 3 5 7
0.69 0.65 0.70 0.68 0.52 0.53 0.66 0.64 0.71 0.76
36.7 39.0 36.2 37.3 22.4 21.6 0.0062 0.0063 0.0057 0.0052
1.86 1.75 1.89 1.83 1.44 1.49 1.61 1.58 1.75 1.92
0.56 0.41 0.54 0.60 0.26 0.26 0.64 0.67 0.87 0.86
41.4 48.3 42.7 40.0 25.3 24.7 0.0057 0.0055 0.0034 0.0035
3.43 0.49 2.28 3.75 0.61 0.004 0.0011 0.0013 0.0001 0.0008
iPLS model SBY SMY k-SBY
1d 2d 2d
3 3 5
0.64 0.52 0.71
39.3 22.2 0.0057
1.83 1.49 1.75
0.53 0.11 0.84
42.7 30.0 0.0038
2.47 1.68 0.0001
285
T.P. Kandel et al. / Bioresource Technology 146 (2013) 282–287
3.2. Comparison of PLS model with previous studies Previous studies (Grieder et al., 2011; Lesteur et al., 2011; Raju et al., 2011) have reported the NIRS based prediction of SMY but they did not report SBY as those studies did not show large difference in CH4 concentrations among samples. The PLS model used to predict SMY in this study was less successful as compared to these previous studies (Table 3). All three previous studies have reported R2 values of calibration and validation models in the range of 0.69– 0.79 which are considerably higher as compared to this study. Similarly, RPD values of NIRS model in this study is slightly lower than the previous studies. Raju et al. (2011) suggested using homogenous samples in batch assays to improve SMY models based on NIRS spectroscopy. Samples used for the batch assay and the NIRS spectra measurements were powdered and homogenized in this
study as suggested but the model performance was not improved. Raju et al. (2011) also suggested using a wider range of SMY values in the reference data set for PLS model calibration to improve SMY prediction. The reference data used in this study had a wide range of chemical composition (Kandel et al., 2013a) as the biomass was harvested regularly from the beginning of the growing season to maturity. As expected, the biomass with lower concentration of structural components produced higher amount of biogas but lower methane concentrations in the biogas produced from the biomass offset the total methane yield. This resulted in a larger range of SBY compared to SMY in the reference data (Table 1). The reference data used in this study had a relatively smaller range of SMY as compared to Lesteur et al. (2011) and Raju et al. (2011) but the range was wider than the data used by Grieder et al. (2011)
0.3
Specific Biogas Yield (SBY) LV1 (98.33) LV2 (0.89) LV3 (0.44)
Loading Weights
0.2
0.1
0.0
-0.1
0.3
Specific Methane Yield (SMY) 0.2 LV1 (96.18%) LV2 (3.11%) LV3 (0.39%)
Loading Weights
0.1
0.0
-0.1
-0.2
-0.3
-0.4
Kinetics of Specific Biogas Yield (k-SBY) 0.2
0.1
Loading Weights
The iPLS models did not improve the models based on full spectrum data for any of the studied parameters (SBY, SMY and k-SBY) (Table 2). When there is high degree of redundancy in NIRS spectra, the iPLS method can improve full-spectrum PLS models as iPLS method eliminates the redundancy (Nørgaard et al., 2000). Therefore, the lower prediction ability of iPLS models in this study probably indicates lower redundancy in the NIRS spectra, meaning that the PLS models use information from large part of the NIRS spectra. As the iPLS models did not improve the model performance, the PLS models based on full spectrum data were used for further comparisons. Loading plots for latent variables are shown in Fig. 2. The loading plot makes it possible to identify the spectral regions that are important for the models. The loading plot (Fig. 2) shows a very complex pattern and that various spectral regions from 1800 to 2500 nm are important. This is also supported by lower prediction ability of the iPLS model compared to the model based on full spectrum data. However, the first latent variable (LV1) which explains >95% variations in Y-data have the highest absolute value around 1800–1900 nm spectral region. Previous studies have shown that the spectral region around 1900 nm describe the lignin content of the biomass (Uner et al., 2009). Similarly, spectral region around 1900 and 1930 are one of the few regions assigned to cellulose (Jones et al., 2006). Previous studies have also shown that chemical composition such as lignin and cellulose concentration of the biomass can predict SBY and SMY (Grieder et al., 2011; Triolo et al., 2011; Kandel et al., 2013a). As the spectral regions, that are important for lignin and cellulose concentration, also explain SBY and SMY to some extent, it is clear that NIRS could directly predict biodegradability of various kinds of biomass. Performance indicator statistics of the PLS models with the best pre-processing methods are also presented in Table 2. The corresponding measured versus predicted plots are presented in Fig. 3. The SBY and k-SBY models were better than the SMY for both the calibration and validation models. The better prediction of SBY by PLS models suggests that the information present in NIRS spectroscopy data explained biodegradability of biomass better but the composition of the final products (concentration of CO2 and CH4) of the digestion was not be explained well. Grieder et al. (2011) also reported that SBY was predicted successfully by NIRS model but the model was less successful in predicting methane concentration dynamics during the batch assay. The reference data used in this study had large difference in proportion of CH4 and CO2 in the biogas which may be difficult to explain by information obtained from the NIRS spectroscopy data. Also, models based on chemical composition of the biomass had predicted SBY better as compared to SMY (Kandel et al., 2013a). The better performance of k-SBY indicates that the rate of degradation of biomass could be predicted well with NIRS model which is important for determination of the feeding rate into the biogas reactor.
0.0
-0.1
LV1 (96.11%) LV2 (3.18%) LV3 (0.38%) LV4 (0.07%) LV5 (0.07%)
-0.2
-0.3
-0.4 1200
1400
1600
1800
2000
2200
2400
Variables Fig. 2. Loading plots for latent variables (LVs). Optimum numbers of LVs were three for SBY and SMY, and five for k-SBY. Variations represented by the latent variables (LVs) are given in percentages inside bracket.
T.P. Kandel et al. / Bioresource Technology 146 (2013) 282–287
680
620
560
500
440
420
0.09
390
0.08 -1
Modelled SMY [NL CH4 (kg VS)-1]
Modelled SBY [NL biogas (kg VS)-1]
740
Modelled k-SBY (day )
286
360 330 300 270
440
500
560
620
680
740
240
Measured SBY [NL biogas (kg VS)-1]
0.06 0.05 0.04
240
380 380
0.07
0.03 270
300
330
360
390
0.03
420
0.04
0.05
0.06
0.07
0.08
0.09
Measured k-SBY (day-1)
Measured SMY [NL CH4 (kg VS)-1]
Fig. 3. Scatter plots of measured vs. predicted for the PLS models. Closed circles represent data from calibration models and open circles represent data from validation models.
Table 3 Comparison of PLRS model performances of specific methane yield (SMY) in this study with previous studies. Study
Calibration model
Validation model
Performance statistics
Grieder et al. (2011) Lesteur et al. (2011) Raju et al. (2011)a This study a
SMY (Reference data)
Performance statistics
SMY (Reference data)
R2
RMSEcv
RPD
Mean
SD
Range
R2
RMSEP
RPD
Mean
SD
Range
0.76 0.79 0.69 0.53
6.2 31 37.4 21.6
2.07 2.13 1.75 1.49
329 234 288 337
13 66 66 32
295–355 23–400 51–406 264–401
0.77 0.76
6.4 28.0
2.08 2.36
329 227
13 57
301–357 87–322
0.26
24.7
1.20
338
29
272–396
Did not present different calibration and validation models.
Table 4 Comparison of different models of specific methane yield (SMY) based on chemical composition and PLS methods. Model
R2
RMSE
Bias
RPD
Lignin ? SMY Cellulose ? SMY NDF ? SMY ADF ? SMY Lignin + Cellulose ? SMY NIRS ? SMY
0.36 0.32 0.42 0.35 0.37 0.53
26.5 27.4 25.2 26.7 26.3 21.6
0.00 0.01 0.01 0.01 0.01 0.00
1.47 1.42 1.54 1.46 1.48 1.49
(Table 3). The reason for lower success of PLS model in this study as compared to the previous studies is not clear. However, as stated earlier, the difference in proportion of CH4 and CO2 in the biogas from the different samples might be one reason for poor performance as the SBY model shows performance similar to SMY models in previous studies (Table 1).
4. Conclusions This study reported NIRS based prediction of specific biogas yield, kinetics of biogas yield, and specific methane yield of RCG biomass. The iPLS models did not improve the models based on full spectrum data. Various spectral regions from 1800 to 2500 nm were important for prediction efficiency indicating the importance of fiber fractions such as cellulose and lignin concentrations. The calibration and validation models for SBY and k-SBY were better than the model for SMY. Although NIRS model presented in this study was less successful compared to the previous studies, the model prediction was better than the models based on chemical composition. Acknowledgements The study was supported by the European Regional Development Fund as a part of the projects ENERCOAST (http://enercoast.net) and BioM (http://www.biom-kask.eu).
3.3. Comparison of the PLS model with models based on chemical composition
References
Comparison of performance of SMY models based on chemical composition of RCG with performance of the PLS model is presented in Table 4. The PLS model showed better performance as compared to the models based on chemical components. A reason for this could be that the NIRS spectra may contain information not only about the composition of the biomass, such as lignin and cellulose content, but it also contains information about associations among such components (Grieder et al., 2011). Such additional information that can be present in NIRS spectra may have contributed to the better NIRS model as compared to the models based on chemical composition. However, the PLS model is still not strong enough to predict the methane yield with a satisfactory precision.
Barnes, R.J., Dhanoa, M.S., Lister, S.J., 1989. Standard normal variate transformation and de-trending of near-infrared diffuse reflectance spectra. Appl. Spectrosc. 43, 772–777. Bruni, E., Jensen, A.P., Pedersen, E.S., Angelidaki, I., 2010. Anaerobic digestion of maize focusing on variety, harvest time and pretreatment. Appl. Energy 87, 2212–2217. Geladi, P., McDougall, D., Martens, H., 1985. Linearization and scatter-correction for near-infrared reflectance spectra of meat. Appl. Spectrosc. 39, 491–500. Gislum, R., Micklander, E., Nielsen, J.P., 2004. Quantification of nitrogen concentration in perennial ryegrass and red fescue using near-infrared reflectance spectroscopy (NIRS) and chemometrics. Field Crops Res. 88, 269– 277. Grieder, C., Mittweg, G., Dhillon, B.S., Montes, J.M., Orsini, E., Melchinger, A.E., 2011. Determination of methane fermentation yield and its kinetics by near infrared spectroscopy and chemical composition in maize. J. Near Infrared Spectrosc. 19, 463–477.
T.P. Kandel et al. / Bioresource Technology 146 (2013) 282–287 Gunaseelan, V.N., 2007. Regression models of ultimate methane yields of fruits and vegetable solid wastes, sorghum and napiergrass on chemical composition. Bioresour. Technol. 98, 1270–1277. Herrmann, A., Rath, J., 2012. Biogas production from maize: current state, challenges, and prospects. 1. Methane yield potential. Bioenergy Res. 5, 1027– 1042. Jones, P.D., Schimleck, L.R., Peter, G.F., Daniels, R.F., Clark III, A., 2006. Nondestructive estimation of wood chemical composition of sections of radial wood strips by diffuse reflectance near infrared spectroscopy. Wood Sci. Technol. 40 (8), 709–720. Kandel, T.P., Elsgaard, L., Lærke, P.E., 2012. Measurement and modelling of CO2 flux from a drained fen peatland cultivated with reed canary grass and spring barley. Global Change Biol. Bioenergy. http://dx.doi.org/10.1111/gcbb.12020. Kandel, T.P., Sutaryo, S., Møller, H.B., Jørgensen, U., Lærke, P.E., 2013a. Chemical composition and methane yield of reed canary grass as influenced by harvesting time and harvest frequency. Bioresour. Technol. 130, 659–666. Kandel, T.P., Elsgaard, L., Karki, S., Lærke, P.E., 2013b. Biomass yield and greenhouse gas emissions from a drained fen peatland cultivated with reed canary grass under different harvest and fertilizer regimes. Bioenergy Res.. http://dx.doi.org/ 10.1007/s12155-013-9316-5. Lesteur, M., Bellon-Maurel, V., Gonzalez, C., Latrille, E., Roger, J.M., Junqua, G., Steyer, J.P., 2010. Alternative methods for determining anaerobic biodegradability: a review. Process Biochem. 45, 431–440. Lesteur, M., Latrille, E., Maurel, V.B., Roger, J.M., Gonzalez, C., Junqua, G., Steyer, J.P., 2011. First step towards a fast analytical method for the determination of biochemical methane potential of solid wastes by near infrared spectroscopy. Bioresour. Technol. 102, 2280–2288. Møller, H.B., Sommer, S.G., Ahring, B.K., 2004. Methane productivity of manure, straw and solid fractions of manure. Biomass Bioenergy 26, 485–495.
287
Nørgaard, L., Saudland, A., Wagner, J., Nielsen, J.P., Munck, L., Engelsen, S.B., 2000. Interval partial least-squares regression (iPLS): a comparative chemometric study with an example from near-infrared spectroscopy. Appl. Spectrosc. 54, 413–419. Raju, C.S., Ward, A.J., Nielsen, L., Møller, H.B., 2011. Comparison of near infra-red spectroscopy, neutral detergent fiber assay and in-vitro organic matter digestibility assay for rapid determination of the biochemical methane potential of meadow grasses. Bioresour. Technol. 102, 7835–7839. Savitzky, A., Golay, M.J.E., 1964. Smoothing and differentiation of data by simplified least squares procedures. Anal. Chem. 36, 1627–1638. Scaglia, B., Confalonieri, R., D’Imporzano, G., Adani, F., 2010. Estimating biogas production of biologically treated municipal solid waste. Bioresour. Technol. 101, 945–952. Seppälä, M., Paavola, T., Lehtomäki, A., Rintala, J., 2009. Biogas production from boreal herbaceous grasses – specific methane yield and methane yield per hectare. Bioresour. Technol. 100, 2952–2958. Shetty, N., Gislum, R., 2011. Quantification of fructan concentration in grasses using NIR spectroscopy and PLSR. Field Crops Res. 120, 31–37. Tilman, D., Hill, J., Lehman, C., 2006. Carbon-negative biofuels from low-input highdiversity grassland biomass. Science 314, 1598–1600. Triolo, J.M., Sommer, S.G., Møller, H.B., Weisbjerg, M.R., Jiang, X.Y., 2011. A new algorithm to characterize biodegradability of biomass during anaerobic digestion: Influence of lignin concentration on methane production potential. Bioresour. Technol. 102, 9395–9402. Uner, B., Karaman, I., Tanriverdi, H., Özdemir, D., 2009. Prediction of lignin and extractive content of Pinus nigra Arnold. var. Pallasiana tree using near infrared spectroscopy and multivariate calibration. J. Wood Chem. Technol. 29 (1), 24–42.