ww.sciencedirect.com

b i om a s s an d b i o e n e r g y 4 8 ( 2 0 1 3 ) 2 2 4e2 3 0

Available online at w

ht tp: / /www.elsevier .com/locate/biombioe

The potential for online monitoring of short-term processdynamics in anaerobic digestion using near-infraredspectroscopy

L. Christian Krapf a,*, Hauke Heuwinkel a, Urs Schmidhalter b, Andreas Gronauer c

a Institute for Agricultural Engineering and Animal Husbandry, Bavarian State Research Center for Agriculture, Voettinger Straße 36, 85354

Freising, GermanybDepartment of Plant Sciences, Technische Universitat Munchen, Emil-Ramann-Straße 2, 85350 Freising, Germanyc Institute of Agricultural Engineering, Department for Sustainable Agricultural Systems, University of Natural Resources and Life Sciences

Vienna, Gregor Mendel Straße 33, 1180 Vienna, Austria

a r t i c l e i n f o

Article history:

Received 10 November 2011

Received in revised form

22 June 2012

Accepted 20 October 2012

Available online 27 December 2012

Keywords:

Biogas production

NIR spectroscopy

PLS regression

Total volatile fatty acids

Volatile solids

* Corresponding author. Tel.: þ49 176 238184E-mail address: krapf.christian@gmail.co

0961-9534/$ e see front matter ª 2012 Elsevhttp://dx.doi.org/10.1016/j.biombioe.2012.10.0

a b s t r a c t

This work reports the development of a near-infrared (NIR) spectroscopy online calibration

for monitoring volatile solids (VS) and total volatile fatty acids (VFA) process parameters

during anaerobic digestion (AD). The objective was to investigate its potential to estimate

the dynamics of AD process parameters after feeding events. A recirculation loop was

employed to record online measurements during an eight-month experiment using

a 3.3 m3 pilot plant fed with maize silage under mesophilic conditions. Sampling was

performed to conduct calibrations and subsequent test-set validations, comparing the NIR

spectroscopy estimates to the reference values. The calibrated accuracy in terms of mean

prediction errors (RMSEP) was 3 g kg�1 for VS and 0.9 g kg�1 for VFA in the fresh matter. By

applying the calibrations to time series spectra, the model accuracy provided adequate

indications of the concentration changes, including highly sensitive monitoring of short-

term VFA dynamics.

ª 2012 Elsevier Ltd. All rights reserved.

1. Introduction

Feeding fluctuations can induce process instabilities in agri-

cultural anaerobic digesters used for producing biogas [1],

particularly when using easily hydrolysable feedstock that is

rich in starch, which can result in the rapid formation of

volatile fatty acids (VFA). The accumulation of these inter-

mediates during the anaerobic digestion (AD) process may be

responsible for a drop in pH after depletion of the buffering

capacity [2,3]. This pH drop significantly slows downmethane

formation, inducing the risk of arresting the AD process.

76; fax: þ49 8161 71 4363m (L.C. Krapf).ier Ltd. All rights reserve27

Therefore, digesters are often operated at a less-than-

optimum loading rate [1]. However, from an economical

point of view, full load operation is demanded as long as the

requirements of high conversion efficiencies are being met.

VFA are usually measured offline via wet chemical analysis.

The time delay between sampling and acquiring the analytical

results is a limiting factor, as it does not allow for real-time

assessment [4]. Considering that acid accumulation can

occur within days if not hours, wet chemistry is not an effi-

cient tool for controlling AD. Therefore, to achieve optimum

loading rates, the process must be monitored online. As

.

d.

b i om a s s a n d b i o e n e r g y 4 8 ( 2 0 1 3 ) 2 2 4e2 3 0 225

a process analytical technology, near-infrared (NIR) spectros-

copy is considered to be a promising method for biological

process control and has shown its potential in the food sector,

animal nutrition, and the chemical industry [5]. Several

studies focussing on NIR technology for the monitoring of

different AD process indicators have been reported recently.

In an offline approach, Holm-Nielsen et al. [6] developed

calibrations for total solids (TS), volatile solids (VS), ammo-

nium, organic nitrogen, and VFA, including single acids, using

slurry from two commercial biogas plants. An online

approach under laboratory conditions was conducted by

Lomborg et al. [7], in which cattle manure that had been

partially mixed with maize silage was used for the calibration

of TS, VS, VFA, and single acids. Jacobi et al. [8] investigated

the long-term applicability for the online estimation of

VFA, acetic, and propionic acids at a large-scale biogas

plant fed with maize silage. In previous work, we investigated

the potential for the development of robust feedstock

calibrations [9].

These studies demonstrate the high potential of spectro-

scopic methods for monitoring agricultural AD processes.

However, further research is required to investigate the

capability of these methods to adequately describe the

dynamics of the AD processes that can occur during biogas

plant operations. The aim of the present study was to inves-

tigate the applicability of NIR spectroscopy for the online

monitoring of concentration changes in VS and VFA after

feeding events. For this purpose, a pilot plant biogas digester

fed with maize silage was operated for eight months under

highly variable feeding conditions. Thus, the NIR calibration

stability for long-term indications of VS and VFA concentra-

tion courses was evaluated; its potential to visualise short-

term changes of these two parameters and their interaction

after feeding events was also examined.

Fig. 1 e Concentrations of volatile solids (VS) and total volatile f

induced by excess feeding. The OLR refers to the amount of VS

matter.

2. Material and methods

2.1. Experimental setup and process operation

A standing tank digester stirred at 10-min intervals by a verti-

cally mounted agitator was used. With a total capacity of

3.3 m3 (height, 270 cm; diameter, 126 cm; liquid volume,

2.5 m3), the digester was operated under mesophilic condi-

tions. For more than six months prior to the experiment, the

digester was fed at an organic loading rate (OLR), measured in

terms of VS, of 2 kgm�3 d�1. Gas productionwasmeasured via

a drum-type gasmeter (Ritter GmbH, Bochum,Germany) at 10-

min intervals. The gas quality (CO2 and CH4) was recorded via

infrared sensors (Awite GmbH, Langenbach, Germany) every

2 h. The digester temperature was regulated by a sensor

located on the sidewall of the digester, which automatically

controlled the flow of heating water through the closed heat-

ing circuit. Changes in temperature (logged at 10-min inter-

vals) over the range 31e47 �C were induced during the

experiment to include the effects that temperature may have

on the spectra. The digesterwas fedwithmaize silage using an

averageVS concentration of 33%of freshmatter (FM). Irregular

feeding was administered to generate strong variations in the

AD process, with the OLR ranging from no feeding to

35 kg m�3 d�1 (Fig. 1), which corresponds to a hydraulic

retention time of nine days.

During feeding, digester slurry was pumped from the lower

part of thedigester intoa tankwhere itwasmixedwith thesilage

beforebeingrepumpedintotheupperregionofthedigesterusing

a spiral pump.When theOLRwas increased, the ratioof silage to

digester slurry was kept constant by continuous repumping of

the mixture into the digester while adding silage into the tank.

Five phases of organic overload were induced by strongly

atty acids (VFA) during the five phases of organic overload

fed to the digester during one feeding event. FM [ fresh

b i om a s s an d b i o e n e r g y 4 8 ( 2 0 1 3 ) 2 2 4e2 3 0226

increasing the amount of silage that was fed to the digester

during the 240 days of the experiment. To allow the digester to

recover from each overload, feeding was usually stopped if the

methaneconcentration inthegasdroppedbelow45%prior to the

next feedingevent. For thefirstphaseoforganicoverload,anOLR

between4.4and6.4kgm�3d�1waschosen. For thesecond, third,

and fourth phases of overload, feeding was increased up to

27 kg m�3 d�1 (day 147), resulting in a VFA concentration of

9 gkg�1 FM.When feedingwas stopped, a sudden recoveryof the

digester was observed, as the accumulated acids were quickly

degraded. For thefifthoverload, anOLRof 35kgm�3 d�1 (day190)

was selected, which caused long-term process imbalances

accompanied by large amounts of propionic acid.

2.2. Spectra acquisition

The spectral data were measured using an external-loop that

was specifically developed to enable online measurements

and process sampling (Anamenter, Hogemann GmbH, Garrel,

Germany). The system was directly connected to the sidewall

of the digester. The inlet was located at the bottom of the

digester and the outlet halfway up the side of the digester. The

digester content was pumped through a stainless metal pipe

(5 cm inner diameter) for the duration of the experiment to

provide continuous loop-circulation. The metal pipe of the

loop was insulated to maintain the temperature of the slurry.

A sapphire window was fitted in an upstream section of the

pipe, and a diode-array spectrometer (X-Three, NIR-Online

GmbH, Walldorf, Germany) was mounted on this window.

Spectra from diffuse reflection were recorded, covering the

400e1750 nm region at a spectral resolution of 10 nm (Si and

InGaAs detectors). The measurement area was 13 mm in

diameter. The raw spectral data captured at intervals of 15 s

were averagedover 3min for onedatapoint.White referencing

was performed automatically at irregular intervals depending

on the collection conditions.

2.3. Sampling and reference analyses

In total, 120 samples were collected. During times of intensive

feeding, sampling was conducted directly before and several

hours after each feeding event. During times of reduced

feeding, sampling was usually performed twice a week.

Primary sampleswere collected via a ball valve in the loop that

was situated after the spectrometer. The spectra that were

Table 1 eDescriptive statistics of the calibration and validation(VFA).

Calibration data set (n ¼ 65)a

Range Mean Median

VS (g kg�1 FM) 53.4e86.5 69.7 69.4

VFAa (g kg�1 FM) 0.2e13.1 5.9 4.8

Temperature (�C) 36.4e47.5 39.9 39.8

FM: Fresh matter; SD: Standard deviation.

a For VFA, calculations for the calibration data set were performed after

n ¼ 64.

recorded over a sampling time of 20 s were averaged and

utilised for subsequent calibration. The volume of the primary

sample ranged from 20 to 30 L depending on the viscosity of

the slurry. From this, a 5-L sub-sample was collected before

a 0.5-L sample was bottled and immediately frozen at �18 �Cuntil the wet chemical analysis. Total solid (TS) concentration

was measured after oven drying at 105 �C to constant mass.

The concentration of volatile solids (VS) was determined after

incineration of the oven dried sample at 550 �C in a preheated

muffle furnace (Heraeus GmbH, Hanau, Germany) for 1 h. The

VFA was measured via distillation (UDK 126 D, Velp, Usmate,

Italy) of the untreated sample and subsequent titration of the

distillate to pH 8.8 using NaOH (Titro Line, SI Analytics, Mainz,

Germany). All of the data presented refer to FM. The analyses

were performed in duplicate. If the difference between the

replicates for one sample was higher than 10%, a third

measurement was recorded. The average value of all repli-

cates determined the reference value that was used for cali-

bration. The standard error of the laboratory analysis (SEL) [10]

was 1.7 g kg�1 for VS and 0.19 g kg�1 FM for VFA.

2.4. Multivariate data analysis

Multivariate data analyses were performed with the log (1/R)

spectra of the 1100e1750 nm region using the Unscrambler 9.8

(Camo, Oslo, Norway). For explorative data analysis, principal

component regression (PCR) [11] was conducted with the

complete data set. Each model that was developed for VS and

VFA was tested with a leave-one-out cross-validation. During

this procedure, each sample in the data set was used for

calibration and validation [12]. By looking for unexplained

variances of X (spectra) and Y (reference), the data set was

probed for outliers. Only those samples that were insuffi-

ciently described by the model and at the same time had

a significant influence were removed. This resulted in the

removal of one sample, which was defined as a Y-outlier for

the VFA. The final calibrations for VS and VFAwere developed

using partial least squares (PLS) regression [13] and test-set

validation. The data set was divided into two sub-sets, one

used for calibration and the other for validation. The samples

obtained during the first and second (day 1e83), and fifth (day

159e240) phases of organic overloadwere used for calibration.

The samples from the third and fourth (day 84e158) phases of

overload were used as a test-set (Fig. 1). Therefore, the cali-

bration and validation sets originated from the same data set

data sets for volatile solids (VS) and total volatile fatty acids

Validation data set (n ¼ 55)

SD Range Mean Median SD

7.2 66.3e91.1 78.5 77.7 6.1

4.5 0.6e8.3 4.2 4.0 2.1

2.3 31.2e40.3 38.8 39.4 1.6

the removal of one sample that was defined as an outlier resulting in

b i om a s s a n d b i o e n e r g y 4 8 ( 2 0 1 3 ) 2 2 4e2 3 0 227

and were only independent with respect to time. In total,

65 (64 for VFA) and 55 samples were used for calibration

and validation, respectively. PLS regression was performed

using the raw spectra. Additionally, a first derivative was

applied to reduce morphological structure in the spectra

[14]. The calibrationswith the smallest rootmean square error

of prediction (RMSEP) [15] were selected for time series

evaluation.

2.5. Time series

The selected calibrations were applied to the continuously

measured spectra during the third and fourth phases of organic

overload. The NIR estimates were compared to the 55 reference

values obtained for each parameter. Whether the NIR spec-

troscopy correctly described the long-term trends in the

concentration of both parameters during the variable feeding

was examined. Subsequently, the suitability of the models to

capture short-term changes was tested by focussing on a more

detailed picture of the VS and VFA dynamics estimated by the

PLSmodel. For thispurpose, twodifferent sectionscharacterised

by a VS feed varying from 1.9 to 8.7 kgm�3 d�1 (days 89e93) and

from 7.5 to 20 kg m�3 d�1 (days 144e148) were selected.

3. Results and discussion

3.1. Reference data set characteristics

Descriptive statistics of the calibration and validation data sets

are shown in Table 1. The VS concentrations in the calibration

set varied from 53 to 87 g kg�1 FM; the VS concentrations in the

validation set slightly exceeded this concentration range.

Compared to the validation set, the samples used for calibra-

tion were spread over a wider range with respect to VFA,

ranging from 0.2 to 13.1 g kg�1 FM. The upper value reflects an

acidified AD process, as it is clearly>3 g kg�1 FM, a value that is

Fig. 2 e PLS regression results of the test-set validation (val). LEF

components calculated from the raw spectra (1100e1750 nm). C

fatty acids (VFA) with five PLS components calculated from the

smoothing points, 2nd degree polynomial). Cal n [ 64 (one outli

prediction (g kgL1 FM), FM [ fresh matter.

potentially acritical limit forADinagricultural biogasplants [6].

As collinearity can reduce model stability, correlations in the

calibrationdata set areof interest. A strong correlation (r¼ 0.99)

between the TS and VS was observed, as they differed only by

ash content. Although the VS and VFA concentrations showed

similar behaviour during the experiments (feeding resulted in

a concentration increase, non-feeding in a decrease), only

a small correlation (r ¼ 0.26) was found between the two

parameters within the calibration data set. In contrast, for the

test-set samples, a high correlation between VS and VFA

(r ¼ 0.86) was found. Within the whole data set, a small corre-

lation (r ¼ 0.20) between the two parameters was found, indi-

cating that the parameters vary independently of each other

over the complete experimental period.

3.2. NIR calibration results

Fig. 2 provides an overview of the PLS regression results for VS

and VFA based on test-set validations. Both parameters were

estimated with acceptable accuracy. A mean estimation error

(RMSEP) of 3 g kg�1 FM for the VS is in accordance with results

from another study that reported a RMSECV of 2.6 g kg�1 FM

[6]. Smaller errors were obtained by Lomborg et al. [7], which

may have been because of a reduction in heterogeneity

resulting from the use of ground maize silage in their exper-

iment. The relatively low coefficient of determination

(r2 ¼ 0.76) for VS in the present work compared to Holm-

Nielsen et al. [6] is presumably because of the small overall

variation, which ranges from 66 to 91 g kg�1 FM. To examine

the robustness of the model with respect to temperature

effects, samples measured at different temperatures in the

calibration set were included [16]. Within the calibration set,

the temperature varied from 36 to 47 �C. As larger errors in the

estimates were not observed for the test-set samples at lower

temperatures, it is assumed that the calibration for VS also

accounted for the temperature effects under moderate

extrapolation conditions. The VFA concentration was

T: Calibration (cal) for the volatile solids (VS) with three PLS

al n [ 65, val n [ 55. RIGHT: Calibration results for volatile

first derivative (1100e1750 nm, Savitsky Golay, three

er removed), val n[ 55. RMSEP[ root mean square error of

Fig. 3 e Long-term validation for the third and the fourth phases of overload. NIR estimates and reference values for volatile

solids (VS) and total volatile fatty acids (VFA) are shown together with the OLR and the temperature profile in the digester.

FM [ fresh matter. See also Fig. 1.

b i om a s s an d b i o e n e r g y 4 8 ( 2 0 1 3 ) 2 2 4e2 3 0228

estimated with a RMSEP of approximately 0.9 g kg�1 FM

(r2 ¼ 0.85) using the first derivative. In the literature, the error

for this parameter ranges between 0.2 and 1.6 g kg�1 FM and

has been reported for varying experimental conditions and

different validation methods [6,7]. Jacobi et al. [8] performed

an online calibration using a digester fed with maize silage

and employing an experimental setup comparable to the one

reported here. With a RMSECV of approximately 0.8 g kg�1 FM

(r2 ¼ 0.94), a similar VFA performance was achieved.

3.3. Validation: long-term

The models for the two parameters with the spectral

measurements during the third and fourth phases of overload

are illustrated in Fig. 3. Owing to feeding, an increase in VS

was observed in the digester, and this is adequately described

by the model. Additionally, a steady decrease was observed in

the concentration level after feeding was temporarily stopped

(day 94). However, at times, the reference value for VS obvi-

ously differed from the estimates. The dominant contribution

to this difference seems to be associated with the sampling

error. On day 103, the reference value showed an increase in

VS concentration, a change that seems unlikely, as feeding

was stopped on day 94. In contrast, the NIR estimates indi-

cated a smooth downward trend, which presumably describes

the continuous biodegradation of the organic matter in the

digester. Constitutional and distributional heterogeneity of

the digester slurry may be considered important factors that

can influence the sampling error [17].

As in the case of VS, the estimates for VFA also followed

the feeding-induced trend of the reference analyses. The

differences between the estimated and experimental values

were less pronounced compared to those observed for VS.

This is probably because the determination of VFA is less

sensitive to the sampling error because the acids are present

in the aqueous phase. Starting on day 123, the estimates

decreased and approached an implausible value of �2 g kg�1

FM, which was quite different from the reference data. At that

time, the temperature in the digester had been reduced to

approximately 31 �C. After a stepwise increase of the

temperature to 40 �C (day 129), the estimated values again

resembled the reference values. This drift in the estimates

indicates that, in contrast to that of the VS, the model for VFA

was sensitive to temperature changes that were not consid-

ered during calibration. However, before and after this period,

the estimates for VFA also sometimes fell below zero. This

finding most likely reflects the random error of the model at

low acid concentrations and may also indicate a weakness in

the calibration owing to an offset. During the validation phase,

a decrease in temperature coincided with low VFA concen-

tration, and no clear distinction between the temperature

effect and model stability at low concentrations could be

elucidated. At VFA concentrations below 1 g kg�1 FM, Jacobi

et al. [8] also reported different behaviour to that of the esti-

mates, and similarly, Reed et al. [18] for VFA concentrations

below 0.3 g L�1 in sewage sludge. Hansson et al. [19] previously

suggested that this behaviour is related to the detection limit

of the NIR method, which was found to be 0.3 g L�1 for pro-

pionic acid. For the data presented here, further research is

required to assess the calibration stability at the lower model

boundary for both temperature and VFA.

3.4. Validation: short-term

The VS and VFA concentrations followed a similar pattern

during the full validation phase and so did the NIR estimates.

To investigate whether the estimation of VFA was indepen-

dent of VS concentration changes, Fig. 4 illustrates a more

detailed time course between days 89e93 and days 144e148.

In the first period, six different feedings ranging from 1.9 to

8.7 kg m�3 d�1 were administered. After each feeding, the

estimated VS displayed an abrupt increase. Within less than

1 h, the VS reached a new level and then steadily declined,

reflecting the conversion of organic matter into biogas. The

Fig. 4 e Short-term validation: NIR estimates of volatile solids (VS) and total volatile fatty acids (VFA) for days 89e93 and

days 144e148. Between two consecutive feeding events, the correlation coefficient (r) for the estimated values of VS and VFA

were calculated. On day 148, plugging of the external-loop pump resulted in temporarily erroneous measurements.

FM [ fresh matter. See also Figs. 1 and 3.

b i om a s s a n d b i o e n e r g y 4 8 ( 2 0 1 3 ) 2 2 4e2 3 0 229

intensity of the feeding had a direct impact on the estimates of

both parameters, and even the second feeding of

1.9 kgm�3 d�1 on day 91 was reflected in the NIR estimates. As

in the case of VS, the VFA estimates also showed a continuous

build up in concentration level after feeding. However, when

the VS concentration was already declining, the estimated

VFA still continued to rise and reached its peak several hours

later. This rise then began to drop comparatively fast, but did

not reach the level prior to feeding. Therefore, owing to the

high overall feeding, a steady accumulation of the acids was

observed over this five-day period. In this regard, the NIR

estimates nicely reflected the expectations of the process

behaviour, both for the quite stable VS parameter and the

highly variable VFA parameter. The high degree of indepen-

dence of the two parameters during the initial hours after

feeding is underlined by their consistently low correlation (r),

which was calculated at each interval between the two

feeding events (Fig. 4). Comparison of the estimated values

with the reference values for both parameters demonstrates

that the relative (precision) as well as the absolute (accuracy)

changes in the concentration levels could mostly be captured

via NIR spectroscopy. Thus, the accumulation of VFA could be

observed even at early stages with 2e3 g kg�1 FM.

The estimates for days 144e148 illustrate the reproduc-

ibility of the results within this work. Owing to a feeding of up

to 20 kg m�3 d�1, both parameters were affected more during

this interval. Unfortunately, a malfunction due to plugging of

the loop-circulation pump occurred on day 148, which

affected the estimates. By comparing the VS concentration

before and after feeding on days 144, 145, and 147, the sharp

increase in the NIR estimates and in the reference values is

less than the increase calculated from the actual amount of VS

fed to the digester. This result is possibly explained by the

formation of a floating layer after these heavy feeding events.

This layer of fresh silage did not enter the loop system as its

inlet was located at the bottom of the digester.

4. Conclusions

Highly reproducible NIRmeasurements can be used for online

indication of short-term dynamics of VS and VFA during AD.

The independence and high sensitivity in the estimation of

both parameters underline that (1) the estimates for VFA were

independent of the VS content and (2) the high potential this

technology offers for theoptimisedprocess operationof biogas

digesters. Themonitoring of relative VS and VFA changesmay

be the key to further process control via NIR spectroscopy, as

feedingmay be adjusted to the current AD process conditions.

b i om a s s an d b i o e n e r g y 4 8 ( 2 0 1 3 ) 2 2 4e2 3 0230

Acknowledgements

This project was financially supported by the Agency for

Renewable Resources (FNR) and the Bavarian State Ministry of

Agriculture and Forests. The authors thank NIR-Online GmbH

and Hogemann GmbH for providing the spectrometer and the

Anamenter.We thank Thomas Schneider and Tomas Qvarfort

for constant support throughout this work.

r e f e r e n c e s

[1] Ward AJ, Hobbs PJ, Holliman PJ, Jones DL. Optimization of theanaerobic digestion of agricultural resources. BioresourTechnol 2008;99(17):7928e40.

[2] Boe K, Batstone DJ, Steyer JP, Angelidaki I. State indicators formonitoring the anaerobic digestion process. Water Res 2010;44(20):5973e80.

[3] Lebuhn M, Liu F, Heuwinkel H, Gronauer A. Biogasproduction from mono-digestion of maize silage e long-termprocess stability and requirements. Water Sci Technol 2008;58(8):1645e51.

[4] Holm-Nielsen JB,DahlCK,EsbensenKH.Representative samplingforprocessanalytical characterizationofheterogeneousbioslurrysystems e a reference study of sampling issues in PAT.Chemometr Intell Lab 2006;83(2):114e26.

[5] Burns DA, Ciurczak EW. Handbook of near-infrared analysis.4th ed. Boca Raton: CRC Press; 2008.

[6] Holm-Nielsen JB, Andree H, Lindorfer H, Esbensen KH.Transflexive embedded near infrared monitoring for keyprocess intermediates in anaerobic digestion/biogasproduction. J Near Infrared Spec 2007;15(2):123e35.

[7] Lomborg CJ, Holm-Nielsen JB, Oleskowicz-Popiel P,Esbensen KH. Near infrared and acoustic chemometricsmonitoring of volatile fatty acids and dry matter during co-digestion of manure and maize silage. Bioresour Technol2009;100(5):1711e9.

[8] Jacobi HF, Moschner CR, Hartung E. Use of near infraredspectroscopy in monitoring of volatile fatty acids inanaerobic digestion. Water Sci Technol 2009;60(2):339e46.

[9] Krapf LC, Gronauer A, Schmidhalter U, Heuwinkel H. Nearinfrared spectroscopy calibrations for the estimation ofprocess parameters of anaerobic digestion of energy cropsand livestock residues. J Near Infrared Spec 2011;19(6):479e93.

[10] Workman Jr JJ. NIR spectroscopy calibration basics. In:Burns DA, Ciurczak EW, editors. Handbook of near-infraredanalysis. Boca Raton: CRC Press; 2008. p. 123e49.

[11] Næs T, Isaksson T, Fearn T, Davies T. Multivariate calibrationand classification. Chichester: NIR publications; 2002.

[12] Martens H, Martens M. Modified Jack-knife estimation ofparameter uncertainty in bilinear modelling by partial leastsquares regression (PLSR). Food Qual Prefer 2000;11(1):5e16.

[13] Wold S, Martens H, Wold H. The multivariate calibrationmethod in chemistry solved by the PLS method. In: Ruhe A,Kagstrom B, editors. Proc. Conf. Matrix Pencils, Lecture Notesin Mathematics. Heidelberg: Springer; 1983. p. 286e93.

[14] Heise HM, Winzen R. Chemometrics in near-infraredspectroscopy. In: Siesler HW, Ozaki Y, Kawata S, Heise HM,editors. Near-infrared spectroscopy. Principles, instruments,applications. Weinheim: WILEY-VCH; 2002. p. 125.

[15] Kessler W. Multivariate Datenanalyse fur die Pharma-, Bio-und Prozessanalytik [Multivariate data analysis for pharma-,bio- and process analytics]. Weinheim: WILEY-VCH; 2002.

[16] Hageman JA, Westerhuis JA, Smilde AK. Temperature robustmultivariate calibration: an overview of methods for dealingwith temperature influences on near infrared spectra. J NearInfrared Spec 2005;13(2):53e62.

[17] Esbensen KH, Geladi P. Principles of proper validation: useand abuse of re-sampling for validation. J Chemometr 2010;24(3):168e87.

[18] Reed JP, Devlin D, Esteves SRR, Dinsdale R, Guwy AJ.Performance parameter prediction for sewage sludgedigesters using reflectance FT-NIR spectroscopy. Water Res2011;45(8):2463e72.

[19] Hansson M, Nordberg A, Sundh I, Mathisen B. Early warningof disturbances in a laboratory-scale MSW biogas process.Water Sci Technol 2002;45(10):255e60.