Bioresource Technology 176 (2015) 233–241

Contents lists available at ScienceDirect

Bioresource Technology
journal homepage: www.elsevier.com/locate/biortech

Early prediction of Biochemical Methane Potential through statistical

and kinetic modelling of initial gas production
Sten Strömberg a,⇑, Mihaela Nistor b, Jing Liu a,b
a
Department of Biotechnology, Lund University, Getingevägen 60, 221 00 Lund, Sweden
b
Bioprocess Control AB, Scheelevägen 22, 223 63 Lund, Sweden

h i g h l i g h t s

 BMP can be predicted at a much earlier stage into a test.

 BMP data from 138 samples divided into five groups are presented.
 61 different prediction algorithms are evaluated for best performance.
 Linear regression of previous test results are important for accurate predictions.

a r t i c l e i n f o a b s t r a c t

Article history: A major drawback of Biochemical Methane Potential (BMP) tests is their long test duration, which could
Received 29 September 2014 be reduced substantially if the final gas production could be predicted at an earlier stage. For this pur-
Received in revised form 6 November 2014 pose, this study evaluates 61 different algorithms for their capability to predict the final BMP and
Accepted 8 November 2014
required degradation time based on data from 138 BMP tests of various substrate types. By combining
Available online 15 November 2014
the best algorithms it was possible to predict the BMP with a relative root mean squared error (rRMSE)
of less than 10% just 6 days after initiation of the experiment. The results from this study indicate that
Keywords:
there is a possibility to shorten the test length substantially by combining laboratory tests and intelligent
Anaerobic digestion
Biochemical Methane Potential
prediction algorithms. Shorter test duration may widen the possible applications for BMP tests in full-
Biogas scale biogas plants, allowing for a better selection and proper pricing of biomass.
Kinetic model Ó 2014 Elsevier Ltd. All rights reserved.
Predicting BMP

1. Introduction
With an increasing demand for efficient waste handling and
fossil fuel replacement, anaerobic digestion has gained a lot of
attention in the recent time. In Europe, the biogas production has
increased substantially in the last few years with a rise from
3.8 Mtoe in 2003 to 12.0 Mtoe in 2012 (EurObserv’ER, 2005,
2013). The growing interest for production of biogas by anaerobic
digestion has led to an increased demand for finding and evaluat-
ing new types of suitable feedstock.
A method to source and determine the feasibility of a material
to serve as a substrate in anaerobic digestion is represented by
the Biochemical Methane Potential (BMP) test. Such a test moni-
tors the gas production following the incubation of an organic
material with an anaerobic bacteria mixture under well-controlled
⇑ Corresponding author. Tel.: +46 222 81 93; fax: +46 222 47 13.
E-mail addresses: sten.stromberg@biotek.lu.se (S. Strömberg), mn@bioprocess-
control.com (M. Nistor), jing.liu@biotek.lu.se (J. Liu).
conditions (Angelidaki et al., 2009). The BMP test provides infor-
mation on the extent of the process and decomposition rate of a
material for biogas production, which are valuable parameters in
the design and operation of a biogas plant (Chynoweth et al.,
1993; Lesteur et al., 2010; Moody et al., 2009; Raposo et al., 2011).
One major drawback with conventional BMP tests is their long
duration. Often 30–60 and sometimes up to 100 days are necessary
to obtain the required results, which means that there is an exten-
sive delay before actions can be taken (Lesteur et al., 2011; Ponsá
et al., 2011). Consequently, the tests are mainly used for pre-stud-
ies, mandatory degradation tests or various scientific applications.
However, if results could be obtained faster, BMP tests would likely
be of interest for industrial biogas producers as well, as this would
allow them to use the test for quality checks and pricing of
substrates.
Previously suggested approaches to obtain quick BMP results
are the use of empirical relationships based on the material’s
chemical and biochemical composition (Thomsen et al., 2014) or
respirometric activity (Scaglia et al., 2010) as well as methods

http://dx.doi.org/10.1016/j.biortech.2014.11.033
0960-8524/Ó 2014 Elsevier Ltd. All rights reserved.
234 S. Strömberg et al. / Bioresource Technology 176 (2015) 233–241
using Near-Infrared Spectroscopy (NIRS) (Doublet et al., 2013). A
more general overview of each technique can be found in the
review article by Lesteur et al. (2010). These approaches show
great potential as they allow immediate results within reasonable
error margins. However, as their predictions are based on time
independent measurements, these techniques do not offer any
information regarding the kinetic degradation of the material.
Moreover, the absence of a biological anaerobic degradation pro-
cess, limits the information about the toxicity and loading rate of
the substrate.
An alternative method to obtain quick BMP results, with the
kinetic and biological degradability information included, is the
combination of conventional BMP tests and numerical prediction
models. Promising results in this regard have been achieved by
Ponsá et al. (2011), observing strong linear correlations between
the biogas production at different times in a test and the final value
for different types of organic fraction of municipal solid wastes
(OFMSW). This suggests that, given a big enough database, it is
possible to statistically predict the final BMP of a sample.
It is commonly known that well-controlled batch degradations
follow certain kinetic patterns that can be modelled using suitable
kinetic models. For complex materials, limited by hydrolytic degra-
dation of particulate matter, a first order rate equation is generally
used to describe the degradation profile (Shahriari et al., 2012).
Several other model types have also been proposed as more accu-
rate alternatives for different types of substrates (Beuvink and
Kogut, 1993; Koch and Drewes, 2014; Rincón et al., 2010). This sug-
gests that, given a reliable set of initial data and an accurate kinetic
model, it is possible to predict the remaining methane potential for
many different types of materials. As this approach includes the
biochemical aspects of anaerobic digestion, key information about
the material’s toxicity and loading strain on the microbial popula-
tion is also obtained. The recent development of specialised labora-
tory equipment, such as Automatic Methane Potential Test System
(AMPTS) II, which allows automatic and reliable gas measurements
with high resolution, makes an approach based on real-time pre-
diction with mathematical models highly feasible.
The current study evaluates 61 different algorithms for their
ability to predict the BMP of various substrate types. Special
emphasis is directed towards finding the most suitable model
and the time required to produce estimations within an acceptable
error range.
2. Methods
2.1. Equipment
The Automatic Methane Potential Test System II (AMPTS II,
Bioprocess Control Sweden AB) was used for the BMP analysis of
the target substrates. The AMPTS II is a standardised laboratory
set-up specially designed for automatic BMP determination of
any biodegradable material. The gas is measured through water
displacement using pre-calibrated flow cells that give a signal for
each 10 mL of produced gas. Temperature and pressure sensors
are used to normalise the gas volume to 0 °C, 1 atm and dry gas
conditions at each measurement point.
2.2. Materials
The BMP was investigated for 138 individual samples divided
into six different groups. In order to describe the experimental var-
iation in a BMP test each replicate (in most cases triplicates) of a
sample was included in the study, making a total of 380 investi-
gated datasets. The studied sample groups were anaerobic sludge
(i.e. effluent from anaerobic digesters), standard compounds (e.g.
cellulose, starch, gelatine), household waste (e.g. fruit and vegetable
waste, milk waste, meat waste, co-digestion mixtures), agricultural
waste (e.g. wheat straw, bamboo waste, banana stem), sewage
sludge (e.g. primary and secondary sludge, co-digestion mixtures)
and lipid rich waste (e.g. butter waste, oil waste). The majority of
the BMP tests were performed using inoculum collected from a
sewage treatment plant (Ellinge sewage plant, Sweden), which
receives municipal wastewater and vegetable residues from the
food industry and operates at mesophilic temperatures.
Sodium hydroxide (reagent grade 97%, Sigma–Aldrich) and thy-
molphthalein pH indicator (dye content 95%, Sigma–Aldrich) were
used for the preparation of 3 M alkaline solution for CO2 fixation.
N2 gas (Air Liquid Gas AB, Sweden) was used to obtain anaerobic
conditions during the sample preparation phase.
2.3. Experimental procedure
The biodegradability of the samples was determined using BMP
assays as described in Strömberg et al. (2014b). The majority of the
tests were performed at 37 °C with an inoculum to substrate ratio
of 2 based on volatile solids (VS) amount. An active volume of
400 ml in 500 ml bottles was used in all tests and mixing was per-
formed mechanically, either intermittently or continuously. Blanks
containing only inoculum were included in every test and used to
deduct the background gas production from the inoculum. No
external nutrients or trace elements were added to the reactors
before starting the BMP tests. All data was sampled with 10 mL
resolution and converted to hourly basis (i.e. one data point per
hour). The BMP (adjusted to 0 °C, 1 atm and dry condition) was cal-
culated by subtracting the gas production of the inoculum from the
gas production of the sample and dividing it by the amount of VS
added.
2.4. Numerical and statistical details
2.4.1. Data structure
In total 380 sample datasets from 138 individual samples,
spread over six groups, were included in the study. The samples
in each sample group were further randomly divided into a calibra-
tion and a validation set containing 67% and 33% of the samples,
respectively. This procedure was repeated five times so that the
samples were represented in both the calibration and validation
set several times. All results presented in this study are based on
the average of these five validation sets.
2.4.2. Statistical parameters
Three statistical parameters (i.e. relative root mean squared
error of prediction (rRMSE), relative absolute error (rAE) and the
coefficient of determination (R2)) were calculated according to
Eqs. (1)–(3):
qP
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi2ffi
^i Þ
ðyi y
n
rRMSE ¼ ð1Þ

y
P jyi y^i j
yi
rAE ¼ ð2Þ
n
P
^ i Þ2
ðy  y
R2 ¼ 1  P i ð3Þ
 Þ2
ðyi  y
^i are the experimental and predicted values,
In Eqs. (1)–(3), yi and y
respectively, for sample i, n is the number of samples and y  is the
mean of the experimental values.
 S. Strömberg et al. / Bioresource Technology 176 (2015) 233–241
2.4.3. Prediction target
The objective of the prediction algorithm was to estimate the
final BMP and time to reach this state. For standardisation pur-
poses, the final BMP was expressed as BMP1%, which corresponds
to the value obtained the first day when the average daily produc-
tion is less than 1% of the total/accumulated BMP (VDI 4630, 2006).
This criterion is necessary as it allows for a standardised way of
expressing final BMP and necessary test time, which otherwise
can be a rather subjective procedure where the reported BMP value
and, especially, the test duration depends on how long time the
process is left to run.
It was expected that the model predictions would improve with
the experiment time as more data is available and the process is
closer to the end. Thus, to cope with this fact, the prediction perfor-
mance was evaluated by measuring the time necessary to reach
two criteria: (i) one based on the time necessary to reach less than
10% in rRMSE and (ii) one based on the time necessary to reach an
R2 higher than 0.9. The criterion for rRMSE was seen as the most
important one as it describes the actual error between experimen-
tal and predicted values. The R2 criterion was included as it indi-
cates how well the algorithm describes the variation in the data.
2.4.4. Numerical calculations
All calculations were performed with MatlabÒ where functions
nlncon and fmincon were used for fitting model parameters for sin-
gle- and multi-variable problems, respectively.
3. Prediction model theory
In this section all evaluated aspects of the prediction algorithm
are described in detail. In order to have a better understanding of
the prediction process, the BMP profiles of three different sub-
strates are presented in Fig. 1 together with the predicted profiles
based on the data collected during the first 3 days of the test.
All evaluated prediction algorithms in this study estimate the
final BMP and required test time of a target sample according to
the following two-step approach: (i) statistical prediction, first a
statistical method is used to estimate a preliminary BMP value
based on a regression analysis of the registered methane produc-
tion from the on-going test and corresponding test data from a
database; (ii) kinetic prediction, the preliminary estimated BMP
value is in a second step used as boundary condition for a simple
kinetic model that predicts the remaining part of the degradation
based on non-linear fitting of the registered methane production
data up until that time point. In general terms, the statistical pre-
diction ensures that a reasonable estimation is provided while the
1000 Household waste
Experimental data
800
BMP (Nml/gVS)
600
400
200
0
0 5
Fig. 1. Accumulated specific methane production for three different samples (i.e. mixed food waste as a household waste, banana stem as an agricultural waste, and butter
residues as a lipid rich waste). The experimental data is represented by the solid line, while the predicted gas profile is showed by the striped line.
235
kinetic prediction acts on the dynamic profile of the gas production
to fine-tune this estimation. As the predictions are continuously
updated for each new measurement, the accuracy is naturally
increasing as the test progress.
3.1. Statistical prediction
The statistical prediction estimates the BMP1% and test time by
cross-referencing the registered BMP profile of the target sample
with the corresponding profiles of the samples in a database. By
utilising similarities in the dynamic profile from previous tests, a
reasonable estimation of the target sample’s BMP profile is
assumed to be achievable. The database consists of data from pre-
viously performed BMP tests and is further organised according to
the sample groups presented in Section 2.2.
For predictions of BMP1% the database samples are used to make
a linear regression, passing through origin, between the produced
BMP during the last half of the test (i.e. BMP(t)  BMP(0.5t);
t = last recorded experimental time of the on-going test) and the
BMP1%. To remove deviating samples, only 10% of the database
samples, based on the smallest rRMSE vs. the target sample, are
used in the regression analysis. In similar fashion, the test time is
predicted by making a linear regression between the time points
when each database sample reach the same relative daily BMP pro-
duction as the target sample and their total test times.
3.2. Kinetic prediction
The kinetic prediction is performed by minimising the differ-
ence between the output from a kinetic model and the gas profile
of the target sample. The idea is to utilise the fact that most degra-
dation profiles are determined by a rate-limiting step, which can
be described by rather simple models. However, as this might
not always be the case, two aspects of the kinetic prediction pro-
cess were investigated in order to find the most favourable option:
(i) kinetic model, different types of kinetic models to cover the var-
ious types of degradation profiles and (ii) data management, differ-
ent options for managing the gas production data.
3.2.1. Kinetic models
In total six different kinetic models (M1–M6) were evaluated for
their accuracy to predict the BMP1% and test time based on the
dynamical gas profile: M1 – a first order rate model (Eq. (4)), M2
– a first order rate model with variable order of time dependency
(Eq. (5)), M3 – a combination of two first order rate models (Eq.
(6)), M4 – a Monod type model (Eq. (7)), M5 – a quadratic Monod
Agricultural waste Lipid rich waste
Predicted data
10 15 20 25
Time (days)
236 S. Strömberg et al. / Bioresource Technology 176 (2015) 233–241

type model (Eq. (8)) and M6 – a modified Gompertz model (Eq. 3.3. Evaluation of the prediction algorithms
(9)):
Three factorial design experiments were used to evaluate the
M1 : BMPðtÞ ¼ BMP1  ð1  expðk1  tÞÞ ð4Þ
impact of the individual model aspect (i.e. kinetic model type, data
management and omission of either the statistical or kinetic pre-
M2 : BMPðtÞ ¼ BMP1  ð1  expðk1  t c ÞÞ ð5Þ diction step). This served as an initial scanning process to identify
which model aspects had a positive respectively negative impact as
M3 : BMPðtÞ well as to learn more about its role in the prediction process. For
¼ BMP1  ð1  X  expðk1  tÞ  ð1  XÞ  expðk2  tÞÞ ð6Þ each investigated aspect, an effect (E) was calculated as the aver-
age difference of a response parameter with and without the model
  aspect included in the prediction algorithm (Eqs. (10)–(12)). The
k1  t
M4 : BMPðtÞ ¼ BMP1  ð7Þ response parameter (Y) was set as the mean value (n = 6) of the
1 þ k1  t
average rRMSE for the predicted BMP1% at 1–2, 3–5, 6–9, 10–15
  and 16–30 days after initiation of the experiment.
t2
M5 : BMPðtÞ ¼ BMP1  ð8Þ Using all possible combinations of a kinetic model, data man-
t 2 þ k1  t þ k2 agement strategy and active or inactive kinetic and statistical pre-
diction step generated a total of 61 possible combinations for the
M6 : BMPðtÞ prediction algorithm, see Table 1.
 
h1  expðk1  tÞ h2  expðk2  tÞ The effects for the three investigated factorial designs were cal-
¼ BMP1  exp   ð9Þ culated according to Eqs. (10)–(12) with the distribution: (i) com-
k1 k2
paring the performance of the prediction with or without the
In Eqs. (4)–(9), BMP1 is the ultimate BMP (NmL/gVS), BMP (t) is the statistical prediction step (Eq. (10)), (ii) comparing the six kinetic
BMP value at time t (day), k1 and k2 (1/day) are rate constants, c (–) models (Eq. (11)) and (iii) comparing the five different variants of
is the order of the time dependence, X (–) is the fraction of readily data management (Eq. (12)).
degradable material and h1 and h2 (1/day) describe the lag period !
for the modified Gompertz model. 1 X X X X
EDB ¼   Y i;j;DB  Y i;j;0 ð10Þ
30 i¼M1:M6 j¼D1:D5 i¼M1:M6 j¼D1:D5
3.2.2. Data management
As the kinetic models predictions often varies depending on !
1 X X X
which part of the gas profile is included in the fitting process, three EM1 ¼  5 Y M1;j;DB  Y i;j;DB ;
25
different alternatives for data selection were evaluated: all – utilis- j¼D1:D5 i¼M2:M6 j¼D1:D5
!
ing all registered data, half – utilising the last half of the registered 1 X X X
data and variable – a variable approach that selects the fraction of EM2 ¼   5 Y M2;j;DB  Y i;j;DB ; . . . ð11Þ
25 j¼D1:D5 i¼M1;M3:M6 j¼D1:D5
data giving the most accurate prediction of the last recorded gas
production. By only utilising data from the later part of the test it !
is possible to avoid negative influence from lag phases and also 1 X X X
ED1 ¼  4 Y i;D1;DB  Y i;j;DB ;
have the model calibrated on the most recent kinetic trend, which, 24 i¼M1:M6 i¼M1:M6 j¼D2:D5
in many cases, should serve to better describe the remaining gas !
1 X X X
production. ED2 ¼  4 Y i;D2;DB  Y i;j;DB ; . . . ð12Þ
The variable procedure is initiated by dividing all registered 24 i¼M1:M6 i¼M1:M6 j¼D1;D3:D5
data into ten different datasets, each including 10% of the data.
The kinetic model is then calibrated with the second last dataset In Eqs. (10)–(12), EDB (%) is the effect for the statistical predic-
and evaluated on how well it predicts the last one (having the most tion step using a database, EM1 and EM2 are the effects for kinetic
recent registered data). This is then compared with the predictions models M1 and M2, ED1 and ED2 are the effects for data manage-
based on data from both the second and third last dataset. If this ment method D1 and D2, Yi,j,DB and Yi,j,0 are the response parame-
expansion of the calibration data improves the prediction, based ters with (DB) and without (0) the statistical prediction step for
on the rRMSE, an additional model calibration is performed involv- kinetic model i and data management method j, see Table 1. For
ing three datasets. This cyclic procedure is repeated until either the Eqs. (11) and (12), an effect is calculated for each kinetic model
expansion of the calibration data leads to a worse fit or all datasets (i.e. M1–M6) and data management strategy (i.e. D1–D5), respec-
are utilised. Finally, the datasets for calibration are combined with tively, by rotating i and j so that the evaluated segment is always
the validation data (last data set) and used to predict the remain- subtracted by all the other segments.
ing, unknown, data.
The kinetic models were also investigated using two different 4. Results and discussion
initial points (i.e. time point when t = 0): original – time zero is
the experimental start time and shifted – time zero is the first data 4.1. Sample characterisation
point used for calibrating the model. A shifted initial point is
included as it makes it possible to utilise the initial dynamic prop- The distribution of BMP1% and degradation time for the different
erties of the model (i.e. model output for small values of t) also at a sample groups are presented with box plots in Fig. 2. As seen,
later stage in the process. anaerobic sludge has the lowest BMP values whereas the lipid rich
In total five combinations of data management strategies were waste has the highest ones. Most variation is observed for house-
investigated: D1 – utilising all registered data with the original hold waste, making it the most representable group for this study.
starting point, D2 – utilising half of the data with the original start- In comparison, the low variation and number of samples for agri-
ing point, D3 – utilising half of the data with a shifted starting cultural waste and lipid rich waste suggests that these sample
point, D4 – utilising the variable data selection approach with groups are less significant for the study. It should be pointed out
the original starting point and D5 – utilising the variable data that even though some points are shown as outliers they are still
selection approach with a shifted starting point. included in the study.
 S. Strömberg et al. / Bioresource Technology 176 (2015) 233–241 237

Table 1
Overview of all evaluated combinations of model aspects for the prediction algorithm.

Statistical prediction Kinetic prediction Indicator No.

Status Status Kinetic model Data handling
Active Active 6 5 DB 30
Active Inactive – – 1
Inactive Active 6 5 0 30
Indicator M1–M6 D1–D5 Total 61

BMP1% 891.3 27.6 Test time 16.2

1000 460.9 (52.1) 40 (4.8) 30.7 (8.5)
(146.7) 27.2 (2.4)
(3.3)
12.1
800 361.3 10.1 (4.4)
(199.8) 30 (3.9)
BMP1% (Nml/gVS)

Test time (days)

12.6
600 (4.1)
360.7 353.1
(26.8) (63.2) 20
224.1
400 (41.4)

73.2 10
200 (21.7)

0 0

Fig. 2. Box plots showing the distribution of BMP1% and necessary degradation time to reach BMP1% for the six investigated sample groups and all samples combined. The
central line corresponds to the median of the group, the top and bottom represents the 75th and 25th percentile, the top and bottom whisker corresponds to the 1.5 times the
difference between the 75th and 25th percentile and, finally, the black dots represents the outliers placed outside of these percentiles. Average values and standard deviations
(in parenthesis) for each sample group are given above each box.

With regard to the test duration, the spread of the data is
greater in comparison to the BMP, with many groups having
around 5 days between the 25th and 75th percentiles. A distinct
separation is observed between anaerobic sludge, agricultural
waste and the lipid rich waste, requiring 20–30 days to reach
BMP1%, and the standard compounds, household waste and sewage
sludge, which mostly require 10–15 days to reach the same level of
biodegradation.
4.2. Evaluation of the prediction algorithms
The results from the factorial design experiments are presented
in Fig. 3 where higher positive values correspond to more accurate
predictions of BMP1%. As seen by the relatively large positive bars
in the bottom left graph, including the statistical prediction step
generally gives more accurate BMP1% predictions compared to
not including it. Only lipid rich waste samples are negatively
impacted with this step activated. Thus, the linear regression based
on the information in the BMP database is important for the perfor-
mance of the prediction algorithm.
A significant difference in the performance of the investigated
kinetic models is observed, with the Monod type (M4), quadratic
Monod type (M5) and first order with variable time dependency
(M2) generally giving positive effects on predictions of BMP1%
while the first order (M1), two combined first order (M3) and the
modified Gompertz (M6) have a negative impact. The three more
successful models allow for a slower decline in gas production in
the later parts of the degradation process, which is beneficial for
samples with high contents of slow degradable complex materials
(Koch and Drewes, 2014). The poor performance of the modified
Gompertz model is somewhat surprising as this model was previ-
ously found to be the most accurate kinetic model in describing the
degradation kinetics of a large set of samples (Strömberg et al.,
2014a). Clearly, there are different requirements for fitting a
kinetic model to a full dynamic BMP profile compared to predicting
the same profile based on an initial fraction of the same data.
With regard to data management, the factorial design results
reveal that a variable approach with a shifted initial point (D5) is
the best alternative for predicting the BMP of the majority of the
substrate groups. Moreover, using half of the data with a shifted
initial point (D3) also provides good results, indicating that a mod-
ified starting point is the most important feature for good predic-
tions. Including all data (D1) has a clear negative effect on the
performance, suggesting that large parts of the early obtained data
compromise the prediction accuracy.
As the factorial design is a method that combines the sum of
several combinations, the results might not reflect the truly best
options for prediction of BMP. This approach should more appro-
priately be used to acquire a general idea on how the individual
options influence the performance of the prediction. The most
successful prediction algorithm combinations are instead better
represented by the lowest value of the response parameter (Y),
which is presented in Table 2.
As seen in Table 2, the statistical prediction step is included for
all samples groups except lipid rich waste, once again demonstrat-
ing the importance of this step for accurate predictions. This is
further strengthened by the fact that the kinetic prediction step is
inactivated in the best performing prediction algorithm for
238 S. Strömberg et al. / Bioresource Technology 176 (2015) 233–241

Anaerobic sludge Standard compounds Household waste

Agricultural waste Sewage slduge Lipid rich waste

3
Kinetic models

1
Effect (%)

-1

-2

-3 M1 M2 M3 M4 M5 M6
First order First order First order Monod type Monod type Modified
variable time two combined quadratic Gompertz
dependency

Statistical
prediction Data management
6 3

4 2
Effect (%)

2 1
Effect (%)

0 0

-2 -1

-4 -2

-6 DB -3
Statistical prediction D1 D2 D3 D4 D5
All Half Half Variable Variable
Original Shifted Original Shifted

Fig. 3. Calculated effects from factorial design experiments evaluating the six different kinetic models (top graph), impact of the statistical prediction step (bottom left graph)
and the five different ways of managing the gas production data (bottom right graph).

Table 2
Best settings for predictions of BMP1% for each sample group based on the response parameter Y, described in Section 3.3.

Sample group Statistical prediction Kinetic model Data management Y (%)

Anaerobic sludge Yes – – 9.9
Standard compounds Yes First order variable time dep. (M2) Variable – shifted (D5) 5.8
Household waste Yes Quadratic Monod type (M5) Variable – shifted (D5) 10.4
Agricultural waste Yes – – 13.5
Sewage sludge Yes – – 4.1
Lipid rich waste No Quadratic Monod type (M5) Half – shifted (D3) 6.3

anaerobic sludge, agricultural waste and sewage sludge. These find-
ings are quite interesting as they imply that reasonable predictions
could be obtained with only the statistical prediction step activated,
which would reduce the computational power demand signifi-
cantly. In fact, the response parameter values (Y) with only the
statistical prediction step activated for standard compounds, house-
hold waste and lipid rich waste are 5.9%, 12.0% and 12.7%, respec-
tively. Thus, with exception for lipid rich waste, the statistical
prediction step performs rather well as a standalone step in the pre-
diction algorithm. As the lipid rich waste group is the smallest
(n = 6) sample group, a better performance by the statistical predic-
tion step is expected if more BMP test data are added to the group.
the initiation of the BMP test vs. the expected BMP1% are presented
in Fig. 4. As is clear from the figure, the model predictions improve
significantly as the experiment progresses. After 1 day, the predic-
tions are rather widely spread, especially for household waste, and
cannot be regarded as accurate. However, already 5 days into the
experiment the predictions are considerably better, shown by the
more uniform distribution of data points along the line. Letting
the test run for 10 days assures rather accurate predictions of
BMP1% for most samples with the exception of lipid rich waste.
In fact, the predictions of lipid rich waste are rather inaccurate
even with 15 days of experimental data and thus, as the predic-
tions are not improved with time, it can be concluded that the
prediction algorithm is not performing well for this sample group.

4.3. Assessment of the prediction accuracy

4.3.2. Overall performance
4.3.1. Experimental results vs. predictions The overall rRMSE, rAE and R2 for the prediction of BMP1% and
By utilising the best prediction algorithm versions, as given in test time are plotted against the duration of a test in Fig. 5. As
Table 2, the predicted BMP1% values after 1, 5, 10 and 15 days from expected, the predictions of both parameters improve with time,
 S. Strömberg et al. / Bioresource Technology 176 (2015) 233–241 239

Anaerobic sludge Standard compunds Household waste

Agricultural waste Sewage sludge Lipid rich waste

1200 1200
Day 1 Day 5
1000 1000
Predicted BMP1% (Nml/gVS)

Predicted BMP1% (Nml/gVS)

800 800

600 600

400 400

200 200

0 0
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200
Experimental BMP1% (Nml/gVS) Experimental BMP1% (Nml/gVS)

1200 1200
Day 10 Day 15
1000 1000
Predicted BMP1% (Nml/gVS)

Predicted BMP1% (Nml/gVS)

800 800

600 600

400 400

200 200

0 0
0 200 400 600 800 1000 1200 0 200 400 600 800 1000 1200
Experimental BMP1% (Nml/gVS) Experimental BMP1% (Nml/gVS)

Fig. 4. Predicted vs. experimental BMP1% for the six studied sample groups at 1, 5, 10 and 15 days from the initiation of the BMP test.

as shown by the decreasing errors and increasing coefficient of
determination along the time axis. With regard to BMP1%, the
R2 > 0.9 criterion is fulfilled already 3 days after starting the test
whereas reaching rRMSE < 10% demands about 6 days. This shows
that the prediction of BMP1% is rather successful on an overall level.
However, with regard to predictions of the test time, the two crite-
ria are met much later into the test (i.e. 18 days for R2 > 0.9 and
25 days for rRMSE < 10%) suggesting that this parameter cannot
be predicted accurately enough, at least not within a reasonable
time. One of the major reasons for this is that the predicted test
time is very dependent on the kinetic parameters and fairly small
errors in their values will propagate to large effects on the pre-
dicted test time. Another factor that makes the prediction of the
test time difficult is the fact that only a small deviation in the mea-
sured gas production might be sufficient to reduce the flow rate
enough to meet the BMP1% criterion and, as a consequence, reach
it earlier than expected. As rate constants are known to vary quite
heavily between different samples and test conditions, accurate
predictions of the experiment time, based on kinetic models, have
to be considered a great challenge.
In Table 3, the time necessary to reach both the R2 > 0.9 and
rRMSE < 10% criteria for each sample group is presented. As seen,
the results vary considerably depending on the type of sample
under investigation. In order to reach the rRMSE criterion, only
1 day is required for the lipid rich waste, whereas 8 days are nec-
essary for the agricultural waste. Most interestingly, the predic-
tions of the most variable group (i.e. household waste) require
only 6 days to reach both criteria, demonstrating that the tech-
nique works for sample groups with high variation in data. It also
suggests that a large number of samples are important for good
prediction accuracy.
The obtained results show that it is possible to predict the BMP
at an earlier stage into the test. However, as the study is performed
under very limited conditions, with mainly one type inoculum and
process temperature, it should primarily be regarded as a proof of
principle and not something widely applicable. In order to have a
full understanding of the process, much more test data, investigat-
ing more conditions, are certainly necessary.
4.4. Comparison with literature data
A summary of the results from previous studies related to pre-
diction of the BMP of an organic material is presented in Table 4. As
seen, most studies with higher rRMSE have lower R2 compared to
the criteria used in this study. In fact, almost all reported values
are reached by the prediction algorithm after 6 days or less. The
only exception is the prediction accuracy achieved by Kandel
et al. (2013), which requires 8 days to reach a lower rRMSE.
240 S. Strömberg et al. / Bioresource Technology 176 (2015) 233–241

40
30
20
10
0 10.6
0.9
0.8
0.7
rRMSE rAE R^2

BMP1% Test time

40 1 40 1

30 0.9 30 0.9
rRMSE and rAE (%)

rRMSE and rAE (%)

R2

R2
20 0.8 20 0.8

10 0.7 10 0.7

0 0.6 0 0.6
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time (days) Time (days)

Fig. 5. rRMSE, rAE and R2 vs. time into the test for predictions of BMP1% and test time of all studied samples. The solid horizontal lines represent a R2 of 0.9 and the striped
horizontal line represents a rRMSE and rAE of 10%.

Table 3
Time required and statistical parameters when rRMSE < 10% and R2 > 0.9 are fulfilled for each sample group.

Sample group rRMSE < 10% R2 > 0.9

2
Time (day) rAE (%) R Time (day) rAE (%) rRMSE (%)
Anaerobic sludge 7 7.7 0.88 9 5.9 7.4
Standard compounds 4 6.7 0.05 15 0.5 2.0
Household waste 6 4.8 0.93 6 4.8 8.6
Agricultural waste 8 5.8 0.75 8 5.8 8.1
Sewage sludge 3 6.5 0.56 6 2.6 4.1
Lipid rich waste 1 7.0 1.51 1 7.0 8.7
All 6 5.4 0.98 3 10.7 17.4

Table 4
Compilation of literature data related to prediction of BMP. Also presented is the time necessary (tPA) by the prediction algorithm to reach the reported rRMSE and R2 for each
study.

Reference Sample type n rRMSE (%) R2 tPA (days)

Biochemical composition
Amon et al. (2007) Maize samples 12 9.4a 0.54 6
Gunaseelan (2007) FVW 24 18.8a 1.25a 3
Gunaseelan, 2009 Solid wastes 32 18.4a 0.02a 3
Triolo et al. (2011) Animal slurries, energy crops, FVW 65 26.5 – 2
Thomsen et al. (2014) Solid wastes 64 19.7 0.96 6
Respirometric activity
Schievano et al. (2008) OFMSW + digestate 46 24.8 0.92 4
Schievano et al. (2009) Solid wastes 81 27.4 – 2
Scaglia et al. (2010) MSW – 23.4b 0.89 3
Near infrared spectroscopy (NIRS)
Lesteur et al. (2011) MSW + lignocellulosic wastes 102 12.3c 0.76 4
Raju et al. (2011) Meadow grasses 95 11.1c,d 0.69 6
Doublet et al., 2013 14 Different types of waste 296 12.6c 0.85 4
Kandel et al. (2013) Reed canary grass 98 7.3c 0.26 8
Triolo et al. (2014) Plant biomass 88 14.5c 0.84 4

Acronyms: FVW – fruit and vegetable waste, MSW – municipal solid waste.
a
Calculated from reported values.
b
Biogas potential of dry matter.
c
Recalculated from average and RMSE.
d
Results from calibration curve.

However, the low R2 reported for this study and the fact that only gated 296 samples divided into 12 groups. As seen in Table 4, the
reed canary grass samples were investigated, implies that the low prediction algorithm presented in this study requires 4 days to
errors might be explained by a lower sample diversity. The most reach similar accuracy for the predictions, which is more than
comparable study is the one by Doublet et al. (2013) that investi- the 2–3 days that is necessary for the sample preparation and
 S. Strömberg et al. / Bioresource Technology 176 (2015) 233–241
NIR analysis used by Doublet et al. (2013). However, it should be
mentioned that information given from the first 4 days of anaero-
bic biological degradation might disclose much information about
possible toxic effects and the rate of the degradation process. Fur-
thermore, as the predictions improve with the experimental time,
the users will always have the option to leave an ongoing experi-
ment to continue longer until their own accuracy requirements
are met.
5. Conclusions
This study demonstrates that mathematical prediction models,
in combination with automated instrumentation, allow for the pre-
diction of BMP for a number of different substrate types. Reason-
able error margins (i.e. rRMSE < 10%) for the prediction of BMP1%
are achieved only 6 days after starting the test. The results also
show that linear regression based on data from previous tests
improves the predictions significantly compared to kinetic model-
ling. However, as this study only evaluates 138 sample types, using
mainly one type of inoculum at mesophilic temperatures, exten-
sive additional testing remains necessary to verify this technical
approach on a wide-scale level.
Acknowledgements
The authors would like to express their gratitude to the Swedish
Energy Agency (2008-000449) and Swedish Research Links pro-
gramme (348-2011-7347) for financial support of this study.
References
Amon, T., Amon, B., Kryvoruchko, V., Zollitsch, W., Mayer, K., Gruber, L., 2007. Biogas
production from maize and dairy cattle manure – influence of biomass
composition on the methane yield. Agric. Ecosyst. Environ. 118, 173–182.
Angelidaki, I., Alves, M., Bolzonella, D., Borzacconi, L., Campos, L., Guwy, A.J.,
Kalyuzhnyi, S., Jenicek, P., van Lier, J.B., 2009. Defining the biomethane potential
(BMP) of solid organic wastes and energy crops: a proposed protocol for batch
assays. Water Sci. Technol. 59 (5), 927–934.
Beuvink, J.M., Kogut, J., 1993. Modeling gas production kinetics of grass silages
incubated with buffered ruminal fluid. J. Anim. Sci. 71, 1041–1046.
Chynoweth, D.P., Turick, C.E., Owens, J.M., Jerger, D.E., Peck, M.W., 1993.
Biochemical methane potential of biomass and waste feedstocks. Biomass
Bioenergy 5 (1), 95–111.
Doublet, J., Boulanger, A., Ponthieux, A., Poitrenaud, M., Cacho Rivero, J.A., 2013.
Predicting the biochemical methane potential of wide range of organic
substrates by near infrared spectroscopy. Bioresour. Technol. 128, 252–258.
EurObserv’ER, 2005. 2005 European barometer of renewable energies. In: 5th
Report, ISBN 2-913620-35-3. Systèmes Solaires, Paris, France.
EurObserv’ER, 2013. The state of renewable energies in Europe, edition 2013. In:
13th EurObserv’ER Report, ISSN 2101–9622. Observ’ER, Paris, France.
Gunaseelan, V.N., 2007. Regression models of ultimate methane yields of fruits and
vegetable solid wastes, sorghum and Napier grass on chemical composition.
Bioresour. Technol. 98, 1270–1277.
Gunaseelan, V.N., 2009. Predicting ultimate methane yields of Jatropha curcas and
Morus indica from their chemical composition. Bioresour. Technol. 100, 3426–
3429.
241
Kandel, T.P., Gislum, R., Jørgensen, U., Lærke, P.E., 2013. Prediction of biogas yield
and its kinetics in reed canary grass using near infrared reflectance
spectroscopy and chemometrics. Bioresour. Technol. 146, 282–287.
Koch, K., Drewes, J.E., 2014. Alternative approach to estimate the hydrolysis rate
constant of particulate material from batch data. Appl. Energy 120, 11–15.
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., Bellon Maurel, V., 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.
Moody, L., Burns, R., Wu-Haan, W., Spajić, R., 2009. Use of biochemical methane
potential (BMP) assays for predicting and enhancing anaerobic digester
performance. In: 44th Croatian & 4th International Symposium on
Agriculture, Opatija, Croatia, 16–20 February, 2009.
Ponsá, S., Gea, T., Sánchez, A., 2011. Short-time estimation of biogas and methane
potentials from municipal solid wastes. J. Chem. Technol. Biotechnol. 86, 1121–
1124.
Raju, C.A., Ward, A.J., Nielsen, L., Møller, H.B., 2011. Comparison of near infra-red
spectroscopy, neutral detergent fibre assay and in-vitro organic matter
digestibility assay for rapid determination of the biochemical methane
potential of meadow grasses. Bioresour. Technol. 102, 7835–7839.
Raposo, F., De La Rubia, M.A., Fernández-Cegrí, V., Borja, R., 2011. Anaerobic
digestion of solid organic substrates in batch mode: an overview relating to
methane yields and experimental procedures. Renew. Sustain. Energy Rev. 16,
861–877.
Rincón, B., Banks, C.J., Heaven, S., 2010. Biochemical methane potential of winter
wheat (Triticum aestivum L.): influence of growth stage and storage practice.
Bioresour. Technol. 101, 8179–8184.
Scaglia, B., Confalonieri, R., D’Imporzano, G., Adani, F., 2010. Estimating biogas
production of biologically treated municipal solid waste. Bioresour. Technol.
101, 945–952.
Schievano, A., Pognani, M., D’Imporzano, G., Adani, F., 2008. Predicting anaerobic
biogasification potential of ingestates and digestates of a full-scale biogas plant
using chemical and biological parameters. Bioresour. Technol. 99, 8112–8117.
Schievano, A., Scaglia, B., D’Imporzano, G., Malagutti, L., Gozzi, A., Adani, F., 2009.
Prediction of biogas potentials using quick laboratory analyses: upgrading
previous models for application to heterogeneous organic matrices. Bioresour.
Technol. 100, 5777–5782.
Shahriari, H., Warith, M., Hamoda, M., Kennedy, K.J., 2012. Anaerobic digestion of
organic fraction of municipal solid waste combining two pretreatment
modalities, high temperature microwave and hydrogen peroxide. Waste
Manag. 32, 41–52.
Strömberg, S., Nistor, M., Liu, J., 2014a. Model evaluation and identification for
improved estimation and data interpretation of BMP tests. In: IWA World
Water Congress & Exhibition, Lisbon, Portugal, 21–26 September 2014.
Strömberg, S., Nistor, M., Liu, J., 2014b. Towards eliminating systematic errors
caused by the experimental conditions in biochemical methane potential (BMP)
tests. Waste Manag. 34, 1939–1948.
Thomsen, S.J., Spliid, H., Østergård, H., 2014. Statistical prediction of biomethane
potentials based on the composition of lignocellulosic biomass. Bioresour.
Technol. 154, 80–86.
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.
Triolo, J.M., Ward, A.J., Pedersen, L., Løkke, M.M., Qu, H., Sommer, S.G., 2014. Near
infrared reflectance spectroscopy (NIRS) for rapid determination of biochemical
methane potential of plant biomass. Appl. Energy 116, 52–57.
VDI 4630, 2006. Fermentation of organic materials – characterisation of the
substrate, sampling, collection of material data, fermentation tests. Verein
Deutscher Ingenieure, Düsseldorf.