You are on page 1of 10

Applied Energy 269 (2020) 115166

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Fuel properties of hydrochar and pyrochar: Prediction and exploration with T


machine learning

Jie Lia, Lanjia Panb, Manu Suvarnaa, Yen Wah Tonga, Xiaonan Wanga,
a
Department of Chemical and Biomolecular Engineering, National University of Singapore, Singapore 117585, Singapore
b
Key Laboratory of Urban Pollutant Conversion, Institute of Urban Environment, Chinese Academy of Sciences, Xiamen 361021, China

H I GH L IG H T S G R A P H I C A L A B S T R A C T

• Datasets for hydrochar and pyrochar


were compiled by systematic litera-
ture review.
• Machine learning models with multi-
task prediction were developed and
optimized.
• Fuel properties of chars were pre-
dicted by optimal models with R of 2

~0.90.
• Feature impacts on targets were ex-
plored with model explainer based on
game theory.
• Performance of models was further
improved after feature re-examina-
tion.

A R T I C LE I N FO A B S T R A C T

Keywords: Conversion of wet organic wastes into renewable energy is a promising way to substitute fossil fuels and avoid
Biochar environmental deterioration. Hydrothermal carbonization and pyrolysis can convert wet organic wastes into
Waste to energy hydrochar and pyrochar, which are potential fossil fuel alternatives due to their comparable fuel properties.
Pyrolysis Machine learning (ML) has strong prediction ability after being trained with historic dataset and facilitates good
Hydrothermal carbonization
understanding of the impact of input features on output targets through a data-driven approach. In this study,
Machine learning
ML models for multi-task prediction of fuel properties of the chars were developed and optimized based on two
Multi-task prediction
datasets for hydrochar and pyrochar. Feature importance and correlation were explored based on optimized ML
model, and feature re-examination was conducted for model improvement. Results showed that support vector
regression model with optimal hyper-parameters exhibited better generalized performance for prediction of both
hydrochar and pyrochar properties with the best average R2 of 0.90 and 0.94. ML-based feature analysis in-
dicated that process temperature and carbon content in the feedstock were the significant features impacting fuel
properties of both chars, while nitrogen content was another important input feature for hydrochar and hy-
drogen content for pyrochar. The accuracy (especially for pyrochar), generalization ability, and computational
speed of models were further improved after feature re-examination. The intuitions obtained from feature
analysis provided meaningful insights to select input features for prediction performance improvement and
computational cost saving, and might guide experiments to produce chars with desired quality.


Corresponding author.
E-mail address: chewxia@nus.edu.sg (X. Wang).

https://doi.org/10.1016/j.apenergy.2020.115166
Received 31 January 2020; Received in revised form 15 April 2020; Accepted 6 May 2020
0306-2619/ © 2020 Elsevier Ltd. All rights reserved.
J. Li, et al. Applied Energy 269 (2020) 115166

1. Introduction amount of data about these processes, and helped researchers to en-
hance their understanding on the existing process and unravel new
The depleting reserves of conventional energy sources, along with insights. In an era of data-centric analytics, machine learning (ML)
their higher rate of consumption than generation have placed an in- serves as a data-driven approach, which can learn exclusively from the
creased demand on the supply of non-conventional energy sources [1]. data without relying on rule-based programs [19,20]. In ML models, the
Various modes of renewable energy have been identified to substitute relevance between input features and output targets are achieved over a
fossil fuels, including solar, wind, geothermal, hydropower and bio- training dataset. This relevance can be further generalized to make a
mass. Among the above listed, biomass, specifically the ones derived prediction of new data samples with the same set of input features [21].
from the waste or residue of wood processing, agriculture and horti- For the ML application on waste to biochar, several works have been
culture are the cheapest to be procured, as well as the most convenient published. Linear regression and random forest models were employed
to be utilized. These kinds of biomass can be recycled and pelletized to to predict the solid yield and the energy recovery of HTC from organic
uniform shape and size for convenient storage, transportation, and wastes with R2 below 0.80 and over 0.90, respectively [22,23]. Two ML
utilization [2,3]. Biomass derived from municipal solid wastes, such as models with high accuracy was developed for biochar yield prediction
food waste, animal manure, sewage sludge etc., are potential but in- from pyrolysis of cattle manure [18]. However, the ML models lacked
tractable sources due to their high organic contents and inherent water generalization ability, as it was developed on a small dataset with only
retention ability [4,5]. Additionally, the large amounts of high-moisture 33 experimental datapoints, obtained in specific lab-scale pyrolysis
organic wastes are considered as carriers of heavy metals, pathogens, process. Couple of other studies proposed the development of two ML
antibiotics, cacosmia, micropollutants, and CO2 emission, thereby algorithms, i.e. adaptive neuro-fuzzy inference system and social-spider
raising serious environmental concerns [6–8]. In such scenario, tech- optimization to predict the biochar yield from the same dataset as
nologies that aim to recover energy and resource from these wet wastes above [24,25]. Although its prediction performance was enhanced, the
provide a solution to avoid environmental deterioration and reduce the generalization of these models was once again limited, given the small
consumption of fossil fuels. size of dataset. Besides, the input features of the above three studies
In the last decade, thermal treatment processes such as hydro- with small datasets only focused on the operational conditions, and did
thermal carbonization (HTC) and pyrolysis have gained significant in- not include any details on feedstock characteristics which are vital to
terest in the research community as potential technologies to recover the char properties. Moreover, in our earlier work, we devised ML
resource and energy from wet wastes, as well as eliminate the organic models to predict the yield and carbon content of biochar derived from
containments, e.g. antibiotics, pathogens [9,10]. The high pressure and dry organic wastes and biomass with improved performance by con-
temperature attained during the hydrothermal process, cause the spe- sidering both feedstock properties and process conditions [20]. How-
cial structures of wet wastes to break down thoroughly, and the sub- ever, it was found that the training and optimization of the ML models,
sequent dewatering process effectively releases the bonded water. For exclusively for predicting each target was a time-consuming and com-
example, the special structures of sludge, e.g. supra colloid with cells putationally expensive process, which was also the universal limitation
and extracellular polymeric substances (EPS), can be disintegrated to in all previous ML studies which primarily developed single-task pre-
release the bonded water under HTC process [11]. It has been proved diction models.
by various researchers that hydrothermal conversion is one of the most Motivated by these abovementioned limitations, in this study, we
promising ways to enhance the dewaterability of sludge and pre-treat aimed to develop ML models to simultaneously predict four different
other wet wastes, due to its high resource recovery potential, high de- fuel properties of hydrochar and pyrochar from high-moisture wastes,
watering efficiency, low cost and easy operation [11,12]. Furthermore, including yield, higher heating value, energy recovery efficiency, and
hydrochar, the solid phase obtained from HTC, is an energy-rich pro- energy densification. Compared to most other studies on ML applica-
duct with high content of carbonaceous material. The hydrochar with tions for biochar, a key novelty of this study was the development of
high calorific values has great potential to be used as a solid fuel to simultaneous multi-task prediction models with good prediction per-
substitute for fossil fuels. More importantly, in the case of HTC, around formance, based on two medium-size datasets containing a compre-
60% of N in the feedstock is transferred to liquid product due to the hensive list of input features. These multi-task ML models tended to be
decomposition of N compounds, which benefits the hydrochar compo- highly efficient, as they required less time and computational costs,
sition and results in lower NOx emission during hydrochar combustion, compared to single-task prediction [26]. To achieve this goal, an ex-
as compared to the raw biomass wastes [12,13]. Alternately, pyrolysis tensive list of publications related hydrochar and pyrochar prepared
is another prevalent technology to convert wet biowaste into energy from HTC and pyrolysis of sludge, food waste and manure were re-
and resource. Multiple valuable products including pyrochar, bio-oil viewed and two separate datasets were compiled for hydrochar and
and biogas are produced during the pyrolysis process [14], and the pyrochar, respectively. The input features containing elementary com-
production of these, depend on the operational conditions. For ex- ponents (C, H, N, O), proximate analysis information (fixed carbon, ash
ample, pyrochar with outstanding physicochemical characteristics can and volatile matter contents) of feedstock, and operational conditions
be produced via slow pyrolysis (300–600 °C with a heating rate of of HTC (temperature, residue time, and water content in reactor) and
1–30 °C/min) [7]. As fuel sources, both pyrochar and hydrochar with pyrolysis (temperature, heating rate and residue time) were taken into
rich carbon content exhibit similar performance to that of coal [15,16], consideration to develop ML models for the prediction of multiple
and possess high energy densification due to the high degree of ar- targets (yield, HHV, ER, and ED). Lastly, the feature re-examination was
omatization [17]. However, not all pyrochar and hydrochar from wet explored to further improve the ML models based on the feature im-
wastes have potential to compete against conventional energy sources, portance of input features and their correlations with these targets
primarily due to the wide range of their caloric values and energy achieving from ML explainer combining with game theory. Given the
densification [7]. Therefore, it is essential to evaluate their fuel prop- impetus garnered by data-driven analytics in recent years, the strategies
erties, such as caloric values and energy densification to determine their and ML models developed in this study can help understand the desired
potential as feasible energy sources. properties of chars to better estimate their application and guide ex-
Conventional methodology to gauge the fuel properties of hydro- periments based on the model inference.
char and pyrochar from wet wastes would be to conduct HTC and
pyrolysis experiments, and thereafter assess the respective fuel char-
acteristics. However, this route is labor-intensive, time-consuming, and
expensive [18]. Substantial publications of lab and pilot-scale work on
HTC and pyrolysis in the recent years, have provided a significant

2
J. Li, et al. Applied Energy 269 (2020) 115166

2. Methodology which is the algorithm’s underlying principle. The types of kernel


functions used in SVM models include linear, sigmoid, polynomial
2.1. Data collection and pre-processing (poly), radial basis function (rbf) and exponential radial basis function
(erbf). This function transforms the input data points (which could be
During the data collection process, a systematic review of literature linear or non-linear in nature), from its original n-dimension space to a
on HTC and pyrolysis processing using sewage sludge, food waste and higher dimensional feature space [28]. This takes into effect by placing
manure as the feed sources was conducted. Published papers on the a linear hyper plane between two types of sample data points, such that
abovementioned topic were thoroughly searched through databases the hyper plane maximizes the margin of data separation.
(e.g. ScienceDirect, ISI Web of Sciences, Google Scholar and RF is a form of ensemble learning method, which works on the
Engineering Village) with keywords such as food waste, sludge, principle of building and combining a multitude of decision trees to
manure, hydrothermal carbonization, pyrolysis and hydrochar etc. A improve its generalization and prediction capabilities [29]. A decision
total of 64 peer-reviewed papers were finally identified, without any tree is non-parametric in nature and consists of two types of nodes,
constraint on the publication date, and from these papers, data was namely the decision nodes and leaf nodes. The decision node evaluates
extracted and compiled. The original datasets are provided in supple- each sample data point on the basis of a test function and passes it on to
mentary material. other branches of the tree depending on its features. During the
The following strategies and assumptions were incorporated during learning process, as the volume of data fed to the decision tree in-
the data collection and pre-processing: creases, the multitude of trees increases - generating hundreds or even
thousands of decision trees [29,30]. The underlying principle in
(1) A total of 248 data points on hydrochar and 165 data points on random forest is the combination of all the decision trees developed
pyrochar were gathered from 38 and 26 published papers with a during learning process, through a process called bagging or bootstrap
total of 24 kinds of sewage sludge, 41 kinds of food waste, and 33 aggregation (Fig. 1b). Each of the individual tree acts as a regressor,
kinds of manure. and the final prediction of the random forest model is derived as the
(2) All the data were impartially accepted, without any authors’ bias mean of the outputs of all the individual decision trees, with minimum
and were meticulously screened to avoid duplicate entries. variance and improved generalization of prediction ability.
(3) The input set of features included feedstock properties determined In this study, both the SVR and RF regression models were devel-
by proximate analysis including fixed carbon (Fc), ash (A), volatile oped using the scikit-learn library in python programming language
matter (V) and element composition (C, H, N, O). All data of these (Python 3.6). The prediction accuracy of SVR is mainly related to the
properties were dry-based values. epsilon (ε), kernel function, and regularization parameter (rp) [31]. The
(4) Parameters including hydrothermal reaction time (Ht), temperature function of the width of ε-insensitive zone, is controlled by epsilon,
(HT), water content (WC), pyrolysis temperature (PT), pyrolysis while fitting the training data. Regularization parameter in the loss
heating rate (PHR), and reaction time at the desire temperature (Pt) function controls the trade-off to achieve low error during the training
comprised of the process related input features. and testing process. Therefore, the hyper-parameters including the
(5) Hydrochar and pyrochar fuel characteristics including yield, higher regularization parameter, epsilon, and kernel function were tuned to
heating value (HHV), energy recovery efficiency (ER) and energy optimize SVR model. The final performance of the RF was the average
densification (ED) were labelled as output targets. The details for prediction error of the trees [32]. For the hyper-parameter tuning of RF,
their calculations are shown in supplementary material. both of the number of trees and the maximum depth of each tree were
(6) The data points were systematically collected from the tables or adjusted from 1 to 20 to achieve the optimized model.
extracted from the figures in the shortlisted literatures.
(7) In order to achieve a uniform range among the variables, all the
data of input features and output targets were normalized using Eq. 2.3. Model training and evaluation of multi-task learning
(1):
xi − μ Machine learning models are classically viewed as an optimization
x i∗ = problem, where the goal is to define a loss function (also known as cost
s (1)
function), which is then minimized so that the model is optimized. The
where x i is the value of input feature i; x i∗ is the normalized value of loss function typically represents the deviation of the predicted values
initial x i ; μ is the mean of x i ; and s is the standard deviation of x i . (derived from the ML model) to that of the actual experimental values.
Lastly, the total data points were divided into two parts. Ninety The loss function of multi-task learning (Ltotal ) as presented in this work
percent of the total data points were randomly selected and labelled as is given by Eq. (2) [33]:
training data, and the remaining data points were labelled as test data
for the final evaluation of the developed models. During the training m

process, the ten-fold cross-validation method was applied to tune the Ltotal (W ) = ∑ λt Lt [h (xit ; wt ), yit ]
t=1 (2)
hyper-parameters of different models for the improvement of their
prediction capability [27]. where m represents the total number of tasks. h is the hypothesis
function of machine learning model. W = [w1,w2 ...wm ] is a matrix made
2.2. Modelling methods and hyper-parameter tuning of parameter vectors (wt ) of hypothesis function learned by minimizing
Ltotal . λt is a scalar coefficient for weighting the importance of loss for
In this work, two state-of-the-art machine learning algorithms, different targets, which is set to be the same for each target in this work.
namely supporting vector machine (SVM) and random forest (RF), were Lt means the loss function of ML models for task t. x it and yit represent
used to predict the fuel characteristics of hydrochar and pyrochar. SVM the values of input and output variables for task t, respectively. i in-
and RF have been widely used in energy domain for prediction of an- dicates the ith data point used in the training process.
nual electricity consumption [27], global incident solar radiation [28], The accuracy of model predictions for both SVR and RF were
and battery capacity [29] etc. evaluated in terms of coefficient of determination (R2) and root mean
A variant of SVM popularly known as support vector regression square error (RMSE). Conceptually, the higher the R2 and lower the
(SVR) was used in this study (Fig. 1a), which suits applications in the RMSE, the greater is the model accuracy [18]. These were calculated as
case of linear and nonlinear regression problems [28]. SVM makes use described in Eqs. (3) and (4):
of an adaptive, margin-based loss function called the kernel function,

3
J. Li, et al. Applied Energy 269 (2020) 115166

Fig. 1. Illustration of (a) Support Vector Regression and (b) Random Forest models.

n t t 2
∑i = 1 (Y pred , i − Yexp, i ) LIME to compute feature importance that can meaningfully interpret
Rt2 = 1 − n t t 2 the outcome of a given machine learning model [34,38].
∑i = 1 (Yexp , i − Yexp, ave ) (3)

n t t 2
∑i = 1 (Y pred , i − Yexp, i ) 3. Results and discussion
RMSEt =
n (4)
3.1. Analysis of input features and output fuel characteristics of chars from
t t
where Y pred , i and Yexp, i are the predicted and the experimental values of wet wastes
t
target t, respectively, while Yexp , ave is the average of all experimental
values of target t. n is the total number of data points and RMSEt re- A descriptive analysis of input features and output targets, derived
presents the root mean squared error of target t. from the compiled datasets is presented in Fig. 2. The data analysis of
feedstock composition including proximate analysis and element com-
2.4. Evaluation of feature impacts on targets position is shown in Fig. 2a. The medians of organic components of C
(carbon) and V (volatile) in dry-based wastes were 42.39% and 73.18%,
Once the ML models are built and their prediction performance respectively, which indicated that alternative fuel with rich C might be
evaluated, interpretation of the models, especially determining feature produced from these types of high-moisture waste. In the case of HTC,
importance is the next gradual step. This is often considered as a HT, Ht, and WC were considered in the operation process, and the
challenging task, given the black-box nature of most ML models. Model medians of these operational parameters were 220 °C and 60 min and
agnostic methods present a possible solution to explain these complex 87.5% respectively (Fig. 2b). Compared to HTC, the median of Pt in
ML models [34]. An extension of the Shapely values from game theory, pyrolysis was the same as Ht, while the PT was 500 °C which was much
SHAP (SHapley Additive exPlanations) is a useful methodology to gain higher than HT (Fig. 2c). Another important operational parameter
insights on the outcome of machine learning predictions [34,35]. The considered in pyrolysis was PHR due to its significant impact on pro-
shapely value is a concept in cooperative game theory to distribute both ducts. The reason for the omitting of WC in pyrolysis was that only a
gains and costs to players/factors, in which two or more players/factors few literatures reported the impact of WC on pyrochar, as most of wet
are involved to achieve the payoff or outcome [36,37]. Therefore this waste feedstocks was dried prior to introduction in the pyrolysis re-
value represents the average marginal contributions of one player/ actor.
factor to the payoff of a strategy with all the combinations and per- During the statistical analysis and inference of the hydrochar and
mutations considered [37]. pyrochar properties, derived from HTC and pyrolysis of wet organic
Conventionally, feature importance in any ML model is determined waste, it was observed that these chars exhibited good fuel properties,
across the entire dataset. In contrast, the SHAP methodology allocates including a similar atomic ratio to coal, high calorific value and good
each feature variable an importance value for every single prediction energy densification. Fig. 2d-f depicts the fuel characteristics of the char
data point, thereby offering local interpretability. The accumulated derived from HTC and pyrolysis of wet wastes. The scatter plot of ER as
SHAP values also determine how much each feature contributes, either a function of yield, along with the violin plots of operational conditions
positively or negatively, to the target variable, thereby making them (i.e. temperature and time) as insets, is shown in Fig. 2d. It is quite
globally interpretable. As reported in literature, calculating the exact evident from the scatter plot that the ER for the hydrochar data points
SHAP value is not a straightforward process [35]. However, by in- exhibited a positive linear correlation with yield, while the pyrochar
tegrating the additive feature attribution method, the SHAP values can data points were slightly scattered. This could be attributed to the wide
be approximated. In this work, we have used the model-agnostic ap- range of operational temperature of pyrochar preparation. Further-
proximation method of Kernel SHAP, which employs a specific type of more, both the yield and ER of hydrochar were higher than that of
weighted linear regression, called the Linear LIME (Local Interpretable pyrochar, thereby indicating that the HTC process could recover more
Model-agnostic Explanations) [38]. Thus, Kernel SHAP is a unified char and retain most of its energy (more than 60%) as compared to
approach which uses Shapely values in conjunction with the Linear pyrolysis. A possible explanation to this phenomenon could arise from

4
J. Li, et al. Applied Energy 269 (2020) 115166

Fig. 2. Statistical data visualization of (a) feedstock composition, (b) operational condition of hydrothermal carbonization, (c) operational condition of pyrolysis, and
fuel characteristics including (d) ER vs. Yield, (e) Van Krevelen diagram, and (f) ED vs. HHV of char from wet wastes (A: Ash; Fc: Fixed carbon; V: Volatile matter; Ht:
Hydrothermal reaction time; HT: Hydrothermal temperature; WC: Water content; PT: Pyrolysis temperature; PHR: Pyrolysis heating rate; Pt: Pyrolysis reaction time;
ER: Energy recovery efficiency; Yield: Mass percentages of hydrochar or pyrochar obtained from wet wastes; ED: energy densification; HHV: Higher heating values).

the fact, that the operational temperatures in the HTC process plots for the hyper-parameter tuning of the ML models. The transition
(~200 °C) were much lower as compared to the pyrolysis process in the color of the contour plot from red to blue indicates model con-
(~500 °C). It is well reported in literature that higher pyrolysis tem- vergence during the hyper-parameter tuning processes. While tuning
perature leads to a greater decomposition of organic components [39], the hyper-parameters of the RF model for the hydrochar dataset, the
and increases energy transfer to the biogas and bio-oil constituents – number of trees = 14 and maximum depth = 8 were identified as
which are the co-products of pyrolysis process in addition to pyrochar. optimized parameters, to obtain a lower RMSE of 0.16. A further in-
The O/C and H/C atomic ratios of organic wastes, hydrochar, and crease in both the number of trees and its maximum depth did not
pyrochar are important parameters to help assess their fuel properties lower the RMSE significantly. Similar trends were observed in the case
for energy application [40]. Literatures stated that chars with low O/C of the RF model developed for the pyrochar dataset. where the number
and H/C atomic ratios possessed high content of aromatic compounds, of trees = 10 and maximum depth = 10 were identified as optimized
stable carbon configuration, and good fuel properties owing to the high parameters to obtain a minimum RMSE of 0.071.
degree of carbonization [39,41]. The O/C and H/C can be visualized in In the case of SVR model, three hyper-parameters, namely, kernel
a Van Krevelen diagram to indicate the chemical reactions during dif- function, regularization parameter and epsilon were tuned for optimi-
ferent thermal treatments, including decarboxylation, dehydration, and zation. For the hydrochar dataset, in the case of linear kernel function,
demethanation, whose vectors are shown in Fig. 2e. According to the the RMSE values decreased until epsilon = 0.5, and then increased with
directions of the above three types of reaction, the decarboxylation the increase of epsilon, but showed no significant decrease in RMSE
reaction was the dominant reactions during the HTC process, while values when the regularization parameter was increased. However, in
both dehydration and demethanation were the main reactions in pyr- the case of poly and radial basis function (rbf) as kernel functions, the
olysis. From the Van Krevelen plot, it can also be inferred that the RMSE values reduced significantly, with a decrease in the epsilon value
atomic ratios of most hydrochar were close to that of peat and lignite, along with simultaneous increase in regularization parameter value.
while most of the pyrochar showed similar atomic ratios with coal and The minimum average RMSE were 0.31, 0.25 and 0.17 for linear, poly
anthracite due to its lower H/C and O/C ratios. These results are in and rbf under tuned regularization parameter and epsilon, respectively.
concordance with the earlier findings reported by Zhou et al. [39]. The above results indicated that rbf function with the regularization
Fig. 2f represents the scatter plot of HHV versus ED, with a nearly parameter of 50 and epsilon of 0.1 exhibited the best prediction per-
positive linear relationship. Most of the hydrochar and pyrochar had formance of hydrochar properties. For the pyrochar prediction model,
HHV and ED values greater than 10 MJ/kg and 1.0, respectively. This the average RMSE values showed a similar trend as described in the
confirmed that the energy in wet wastes was condensed by both the above-mentioned hydrochar model, during the hyper-parameter tuning
HTC and pyrolysis process. On comparing both the chars, the pyrochar process (Fig. 3b). The final minimum average RMSE of 0.062 was ob-
had a steeper slope for ED as a function of HHV, suggesting a faster tained under the optimal condition of rbf function with the regular-
increasing of ED with the increasing of HHV as compared to hydrochar. ization parameter and epsilon to be 10 and 0.1, respectively.

3.2. Model development and hyper-parameter tuning 3.3. Model evaluation

To this aim, two ML models (RF and SVR) were developed and their Once the ML models were developed and optimized as described in
hyper-parameters were tuned based on the ten-fold cross-validated- Section 3.2, they were evaluated for the robustness of their multi-task
average RMSE to predict of hydrochar and pyrochar properties, based prediction capability on the test dataset. The original experimental data
on the dataset described in Section 2.1. Fig. 3 represents the contour versus the predicted values of the hydrochar and pyrochar are

5
J. Li, et al. Applied Energy 269 (2020) 115166

Fig. 3. Hyper-parameter optimization of machine learning models for energy properties prediction of (a) hydrochar and (b) pyrochar (RF: Random forest; SVR:
Support vector regression).

Fig. 4. Multi-task predicted data versus the original experimental data of (a) hydrochar and (b) pyrochar based on the optimized machine learning models with
training and testing datasets (RF: Random forest; SVR: Support vector regression).

Fig. 5. Plots of predicted data versus original experimental data for each target of (a) hydrochar and (b) pyrochar with histograms of datasets at the top and right of
each sub-figure (Yield: Mass percentage of hydrochar or pyrochar obtained from wet wastes; ER: Energy recovery efficiency; HHV: Higher heating values; ED: Energy
densification; RF: Random forest; SVR: Support vector regression).

presented in the form of scatter plots in Fig. 4. Specifically, Fig. 4a compared to the average R2 (0.94 and 0.79) of the RF prediction. This
represents a comparative evaluation between the RF and SVR models to result suggested a better and more robust multi-task prediction per-
predict the fuel characteristics of the hydrochar. It is evident from the formance of SVR due to its higher prediction accuracy of both the
plot that, both the training and test data points predicted by the SVR training and test datasets. Furthermore, the details about the prediction
model were denser than RF. Also, the average R2 for the four targets performance for each fuel characteristic of hydrochar and pyrochar
(based on training and testing datasets) derived from the SVR model with the histograms of datasets at the top and right of each sub-figure
were found to be 0.96 and 0.91, both of which were greater as are shown in Fig. 5. The average R2 of the SVR model for yield, HHV,

6
J. Li, et al. Applied Energy 269 (2020) 115166

ER, and ED of hydrochar in the test dataset were 0.88, 0.96, 0.92, and contribution of every input data point on the targets was quantified as
0.89, respectively (Fig. 5a). SHAP value. A positive SHAP value indicates that the input feature
For the pyrochar dataset (Fig. 4b), both the RF and SVR models had contributes positively to the predicted value. A point of red color re-
the same average R2 (0.97) and similar RMSE for the training dataset. presents a higher original value of the input feature, while blue for
However, the average R2 of RF and SVR model were found to be 0.75 lower value. If the color of the data points belonging to an input feature
and 0.88, respectively, in the test dataset. The above result indicated turns from blue to red with the increase of SHAP value, it is inferred
that SVR facilitated a better prediction performance than RF. Further- that the input feature has a positive impact on the output target. Fur-
more, according to the plot of predicted vs. actual data for each target thermore, Fig. 6c and d depict the stacks of mean absolute SHAP values
in Fig. 5b, the prediction of SVR for the yield, HHV, ER, and ED of of each input for the four characteristics of hydrochar and pyrochar. A
pyrochar in the test dataset had R2 of 0.88, 0.95, 0.79, and 0.91, re- higher sum of mean absolute SHAP values for the four outputs indicates
spectively. It was intrinsic that the prediction performance of ML a greater impact during the multi-task prediction process.
models was lower in test dataset as compared to the training dataset. As depicted in Fig. 6a, the impacts of inputs on various targets were
However, the predicted performance of optimized SVR models was different. For the yield prediction, N content, temperature of HTC (HT)
sufficient and comparable to exhibit a good generalization of fuel and C contents were identified as the three most significant variables,
properties prediction for both hydrochar and pyrochar. In this study, it where both N content of the wet wastes and HT had a negative impact
was evident that the SVR model outperformed the RF model in both the on the yield of the hydrochar. The C content showed no obvious
cases of hydrochar and pyrochar prediction studies. monotonicity with yield, and either a low or high C content could be
associated with high yield. As for the HHV prediction, the HT, C and H
contents were the top three impactful inputs features, and each of them
3.4. Impacts of input features on the fuel characteristics of the char had a positive impact on the HHV of hydrochar. In the case of ER
prediction, N, C and water contents emerged as the three most im-
In order to determine the impacts of the input variables, including portant input features, where the N content was negatively correlated
the elementary and proximate analysis data of the wet wastes and the to ER, while both C and WC contents exhibited a non-monotonicity.
operational conditions of HTC and pyrolysis, on the fuel characteristics Lastly, for ED prediction, the top three positive features were HT, O and
of hydrochar and pyrochar, the ML model explanation was investigated Fc contents. Furthermore, the HT, N and C contents were determined as
using a game theory approach namely SHAP. This analysis was only top three important features for the overall impact on the four targets
conducted for the SVR model which had outperformed the RF model in investigated in the hydrochar dataset (Fig. 6c). It was reasonable that
this study. The impacts of each input feature on the respective output HT and C content were identified as significant features because HT
target of hydrochar and pyrochar are illustrated in Fig. 6a and b. The

Fig. 6. Impacts of input features on each individual characteristic and overall characteristics of hydrochar (a and c) and pyrochar (b and d) based on the explanation
of SVR model using SHAP values.

7
J. Li, et al. Applied Energy 269 (2020) 115166

Table 1
The comparison of prediction performance of fuel properties belonging to hydrochar and biochar ML models developed on the basis of all features and selected
features.
Test performance Hydrochar Pyrochar

R2 RMSE R2 RMSE

All features Updated feature All features Updated feature All features Updated feature All features Updated feature

Yield (%) 0.88 0.86 7.83 8.49 0.88 0.97 5.86 3.36
HHV (MJ/kg) 0.96 0.95 1.39 1.43 0.95 0.96 1.37 1.38
ER (%) 0.92 0.90 6.22 6.86 0.79 0.92 9.40 6.84
ED 0.89 0.89 0.09 0.09 0.88 0.93 0.09 0.10
Average 0.91 0.90 3.89 4.22 0.88 0.94 4.18 2.92

affected the degree of thermal degradation of wastes, and C was the categories of input features namely elementary, proximate and opera-
main source of energy [42,43]. However, another interesting finding in tional conditions with a total of ten input features. Actually, it was
this study was that the N content was also one of the critical factors for reported that not all the input variables in a ML framework contribute
the energy properties of hydrochar, especially for the prediction of yield to the model building process [47]. A similar observation was identified
and ER. In general, the importance of elementary compositions ac- in this study, during the SHAP value analysis as seen Fig. 6. Some
counted for 48% of the four features together (Fig. 6c), indicating that features might be redundant to add to the computational cost of ML
this feature category was the foremost to fuel properties of hydrochar. models, without having a significant impact on the prediction perfor-
Compared to hydrochar, the impact of input features on fuel prop- mance. In the case of hydrochar, it was found that the Ht was the least
erties of pyrochar showed subtle yet comparable differences. In detail, important feature in the determination its four fuel properties, which is
temperature (PT), heating rate (PHR) of pyrolysis, and C content of evidently seen in Fig. 6a and 6c. Therefore, it was removed from the
waste were identified as the three most significant variables impacting model to improve the model generalization ability and reduce the
the yield prediction of pyrochar and were negatively correlated as seen computational cost. Similarly, the three features including Pt, V, and N
in Fig. 6b. Heating rate was a rare and distinct feature impacting pyr- content in the pyrolysis process were discarded due to their negligible
olysis yield, which was not investigated in the case of HTC. From a significance in the model prediction, as determined by SHAP values
process perspective, this can be well understood from the fact that HTC (Fig. 6d). For the purpose of this study, the process of removing re-
operates in lower temperature regimes (~200 °C), whereas the different dundant variables based on the inference of SHAP was termed as
types of pyrolysis (fast or slow) with a higher temperature (~500 °C) “feature re-examination”. Post feature the re-examination, the SVR
are heavily dependent on its heating rate [44]. It is well cited in lit- models were re-devised with the same hyper-parameters as described in
erature that a lower heating rate i.e. slow pyrolysis, yields char as the Section 3.2, but only included the updated input features.
main product, whereas fast pyrolysis yields more of bio-oil [44]. Thus, The comparison of prediction performance of the ML models based
it is proved that the findings in this work are in good agreement with on initial set of features and post feature re-examination in the test
the ones published in literature. For HHV of the pyrochar, the tem- dataset is presented in Table 1. The results showed that the average R2
perature and C content were the two important positively correlated of the ML model post feature re-examination for fuel properties pre-
features, following a similar trend to that of HTC. However, the ash diction of hydrochar decreased marginally from 0.91 to 0.90 with the
content was standout among the top three features for pyrochar in slight reduction of R2 specifically for yield, HHV, and ER prediction.
terms of HHV, in contrast to the H content for HTC. This may be at- Although its prediction accuracy marginally reduced after feature re-
tributed to the fact that the H content in pyrochar is lower than that in examination, the generalization ability and computation process could
hydrochar due to a higher degree of dehydration and demethanation in be improved. In the pyrolysis process, the performance of fuel proper-
pyrolysis [45], while most of the ash remains intact in the pyrochar, ties prediction for pyrochar, significantly increased from an average R2
and this in turn could negatively impact the HHV of the pyrochar [46]. of 0.88 to 0.94 with a much smaller average RMSE value. In detail, the
In terms of ER in pyrolysis, the top three important features were PT, H, R2 of yield, ER, and ED increased from 0.88, 0.79, and 0.88 to 0.97,
and O contents, all of which were markedly different compared to the 0.92, and 0.93, respectively. The above results suggested that the fea-
important features in HTC. The PT and H content had negative influ- ture re-examination based on SHAP values provided useful insights on
ence on the ER, whereas the O content was positively correlated. In the the selection of the input features, and thereby aided the improvement
case of ED prediction of pyrochar, the top three features identified were of prediction performance, model generalization ability, and compu-
O, H and ash contents. The O content was negatively correlated with tational process. The inference drawn from the feature re-examination
ED, whereas the other two had a positive influence. The differences in process can also aid towards an informed data collection process, from a
the impact of input features on targets variables of the chars from HTC futuristic perspective.
and pyrolysis might be attributed to their difference in the degree of
thermal decomposition of feedstocks, mainly due to the gap of their
operational temperatures. In summary, the PT, C and H contents were 4. Conclusions
determined as the three most important input parameters impacting the
four target variables investigated under the pyrochar dataset, and ele- Two types of machine learning models with multi-task prediction
mentary composition was again identified as the most important fea- capability were developed and optimized based on two datasets, com-
ture category to determine the fuel properties of pyrochar, accounting prising of 248 kinds of hydrochar and 166 kinds of pyrochar, respec-
for 46% of importance (Fig. 6d). tively. Support vector regression model with optimal hyper-parameters
showed better prediction performance for both datasets as compared to
random forest model. The final average R2 of the support vector re-
3.5. Feature re-examination for model improvement gression model with feature re-examination were 0.90 and 0.94 for the
four target variables of the hydrochar and pyrochar, respectively.
The ML models developed for fuel properties prediction of hydro- Moreover, the impacts of input features on the properties of char were
char and pyrochar, as described in Sections 3.2 and 3.3, included three explored based on the optimized model. Process temperature and C

8
J. Li, et al. Applied Energy 269 (2020) 115166

content of waste were two significant features that were ranked con- critical review from technological development and environmental impact analysis
sistently high for most of the fuel properties (target variables) for both perspective. Appl Energy 2019:256. https://doi.org/10.1016/j.apenergy.2019.
113961.
the chars. N and H contents in the feedstock were also found to have [9] Li J, Yu G, Xie S, Pan L, Li C, You F, et al. Immobilization of heavy metals in
significant impact the hydrochar and pyrochar, respectively. The pre- ceramsite produced from sewage sludge biochar. Sci Total Environ
diction accuracy (especially for pyrochar), generalization ability and 2018;628–629:131–40. https://doi.org/10.1016/j.scitotenv.2018.02.036.
[10] Li J, Pan L, Yu G, Xie S, Li C, Lai D, et al. The synthesis of heterogeneous Fenton-like
computational speed of models were further improved by discarding catalyst using sewage sludge biochar and its application for ciprofloxacin de-
redundant input variables based on the results of the ML-based feature gradation. Sci Total Environ 2019;654:1284–92. https://doi.org/10.1016/j.
analysis. Lastly, limited datasets and lack of experimental validation of scitotenv.2018.11.013.
[11] Li C, Wang X, Zhang G, Li J, Li Z, Yu G, et al. A process combining hydrothermal
the model predictions were the two limitations of the present work, pretreatment, anaerobic digestion and pyrolysis for sewage sludge dewatering and
which will be investigated in our future studies. co-production of biogas and biochar: Pilot-scale verification. Bioresour Technol
2018;254:187–93. https://doi.org/10.1016/j.biortech.2018.01.045.
[12] Yu Y, Lei Z, Yang X, Yang X, Huang W, Shimizu K, et al. Hydrothermal carboni-
CRediT authorship contribution statement
zation of anaerobic granular sludge: Effect of process temperature on nutrients
availability and energy gain from produced hydrochar. Appl Energy
Jie Li: Conceptualization, Methodology, Formal analysis, Writing - 2018;229:88–95. https://doi.org/10.1016/j.apenergy.2018.07.088.
review & editing, Visualization, Data curation. Lanjia Pan: Formal [13] Zhao P, Chen H, Ge S, Yoshikawa K. Effect of the hydrothermal pretreatment for the
reduction of NO emission from sewage sludge combustion. Appl Energy
analysis, Data curation. Manu Suvarna: Methodology, Formal analysis, 2013;111:199–205. https://doi.org/10.1016/j.apenergy.2013.05.029.
Writing - review & editing. Yen Wah Tong: Writing - review & editing. [14] Laird DA, Brown RC, Amonette JE, Lehmann J. Review of the pyrolysis platform for
Xiaonan Wang: Conceptualization, Writing - review & editing, coproducing bio-oil and biochar. Biofuels, Bioprod Biorefining 2009;3:547–62.
https://doi.org/10.1002/bbb.169.
Supervision. [15] Huang YF, Cheng PH, Te Chiueh P, Lo SL. Leucaena biochar produced by micro-
wave torrefaction: Fuel properties and energy efficiency. Appl Energy
Declaration of Competing Interest 2017;204:1018–25. https://doi.org/10.1016/j.apenergy.2017.03.007.
[16] Ismail TM, Yoshikawa K, Sherif H, Abd El-Salam M. Hydrothermal treatment of
municipal solid waste into coal in a commercial Plant: Numerical assessment of
The authors declare that they have no known competing financial process parameters. Appl Energy 2019;250:653–64. https://doi.org/10.1016/j.
apenergy.2019.05.042.
interests or personal relationships that could have appeared to influ-
[17] Ciuta S, Patuzzi F, Baratieri M, Castaldi MJ. Enthalpy changes during pyrolysis of
ence the work reported in this paper. biomass: Interpretation of intraparticle gas sampling. Appl Energy
2018;228:1985–93. https://doi.org/10.1016/j.apenergy.2018.07.061.
[18] Cao H, Xin Y, Yuan Q. Prediction of biochar yield from cattle manure pyrolysis via
Acknowledgement
least squares support vector machine intelligent approach. Bioresour Technol
2016;202:158–64. https://doi.org/10.1016/j.biortech.2015.12.024.
This work was supported by the National Research Foundation, [19] Zhu X, Wang X, Ok YS. The application of machine learning methods for prediction
Prime Minister’s Office, Singapore under its Campus for Research of metal sorption onto biochars. J Hazard Mater 2019:378. https://doi.org/10.
1016/j.jhazmat.2019.06.004.
Excellence and Technological Enterprise (CREATE) program (Grant [20] Zhu X, Li Y, Wang X. Machine learning prediction of biochar yield and carbon
Number R-706-000-103-281 and R-706-001-102-281). The authors contents in biochar based on biomass characteristics and pyrolysis conditions.
acknowledge the Singapore RIE2020 Advanced Manufacturing and Bioresour Technol 2019:288. https://doi.org/10.1016/j.biortech.2019.121527.
[21] Wu G, Kechavarzi C, Li X, Wu S, Pollard SJT, Sui H, et al. Machine learning models
Engineering (AME) Programmatic grant “Accelerated Materials for predicting PAHs bioavailability in compost amended soils. Chem Eng J
Development for Manufacturing” by the Agency for Science, 2013;223:747–54. https://doi.org/10.1016/j.cej.2013.02.122.
Technology and Research under Grant No. A1898b0043 and the IAF-PP [22] Li L, Flora JRV, Caicedo JM, Berge ND. Investigating the role of feedstock properties
and process conditions on products formed during the hydrothermal carbonization
grant “Cyber-physical production system (CPPS) towards contextual of organics using regression techniques. Bioresour Technol 2015;187:263–74.
and intelligent response” by the Agency for Science, Technology and https://doi.org/10.1016/j.biortech.2015.03.054.
Research under Grant No. A19C1a0018. [23] Li L, Flora JRV, Berge ND. Predictions of energy recovery from hydrochar generated
from the hydrothermal carbonization of organic wastes. Renew Energy
2020;145:1883–9. https://doi.org/10.1016/j.renene.2019.07.103.
Appendix A. Supplementary material [24] Ewees AA, El Aziz MA, Elhoseny M. Social-spider optimization algorithm for im-
proving ANFIS to predict biochar yield. In: 8th Int. Conf. Comput. Commun. Netw.
Technol. ICCCNT 2017, Institute of Electrical and Electronics Engineers Inc.; 2017.
Supplementary data to this article can be found online at https://
https://doi.org/10.1109/ICCCNT.2017.8203950.
doi.org/10.1016/j.apenergy.2020.115166. [25] Aziz MA El, Hemdan AM, Ewees AA, Elhoseny M, Shehab A, Hassanien AE, et al.
Prediction of biochar yield using adaptive neuro-fuzzy inference system with par-
References ticle swarm optimization. In: Proc. - 2017 IEEE PES-IAS PowerAfrica Conf.
Harnessing Energy, Inf. Commun. Technol. Afford. Electrif. Africa, PowerAfrica
2017, Institute of Electrical and Electronics Engineers Inc.; 2017, p. 115–20.
[1] Zendehboudi A, Baseer MA, Saidur R. Application of support vector machine models https://doi.org/10.1109/PowerAfrica.2017.7991209.
for forecasting solar and wind energy resources: A review. J Clean Prod [26] Wang L, Li Q, Chen X, Li S. Multi-task learning for gender and age prediction on
2018;199:272–85. https://doi.org/10.1016/j.jclepro.2018.07.164. Chinese microblog. Lect Notes Comput Sci (Including Subser Lect Notes Artif Intell
[2] Li H, Liu X, Legros R, Bi XT, Jim Lim C, Sokhansanj S. Pelletization of torrefied Lect Notes Bioinformatics) 2016;10102:189–200. https://doi.org/10.1007/978-3-
sawdust and properties of torrefied pellets. Appl Energy 2012;93:680–5. https:// 319-50496-4_16.
doi.org/10.1016/j.apenergy.2012.01.002. [27] Payam R, Tang L, Liu H. Cross-Validation. Encycl. Database Syst., Boston, MA:
[3] Liu Z, Quek A, Balasubramanian R. Preparation and characterization of fuel pellets Springer US; 2009, p. 532–8. https://doi.org/10.1007/978-0-387-39940-9_565.
from woody biomass, agro-residues and their corresponding hydrochars. Appl [28] Panda AK, Rapur JS, Tiwari R. Prediction of flow blockages and impending cavi-
Energy 2014;113:1315–22. https://doi.org/10.1016/j.apenergy.2013.08.087. tation in centrifugal pumps using Support Vector Machine (SVM) algorithms based
[4] Xie S, Yu G, Li C, You F, Li J, Tian R, et al. Dewaterability enhancement and heavy on vibration measurements. Meas J Int Meas Confed 2018;130:44–56. https://doi.
metals immobilization by pig manure biochar addition during hydrothermal treat- org/10.1016/j.measurement.2018.07.092.
ment of sewage sludge. Environ Sci Pollut Res 2019;26:16537–47. https://doi.org/ [29] Prasad AM, Iverson LR, Liaw A. Newer classification and regression tree techniques:
10.1007/s11356-019-04961-1. Bagging and random forests for ecological prediction. Ecosystems 2006;9:181–99.
[5] Diggelman C, Ham RK. Household food waste to wastewater or to solid waste? That https://doi.org/10.1007/s10021-005-0054-1.
is the question. Waste Manag Res 2003;21:501–14. https://doi.org/10.1177/ [30] Zhang S, Tan Z, Liu J, Xu Z, Du Z. Determination of the food dye indigotine in cream
0734242X0302100603. by near-infrared spectroscopy technology combined with random forest model.
[6] Leng L, Yuan X, Shao J, Huang H, Wang H, Li H, et al. Study on demetalization of Spectrochim Acta - Part A Mol Biomol Spectrosc 2019. https://doi.org/10.1016/j.
sewage sludge by sequential extraction before liquefaction for the production of saa.2019.117551.
cleaner bio-oil and bio-char. Bioresour Technol 2016;200:320–7. https://doi.org/ [31] Shataee S, Kalbi S, Fallah A, Pelz D. Forest attribute imputation using machine-
10.1016/j.biortech.2015.10.040. learning methods and ASTER data: Comparison of k-NN, SVR and random forest
[7] Mau V, Gross A. Energy conversion and gas emissions from production and com- regression algorithms. Int J Remote Sens 2012;33:6254–80. https://doi.org/10.
bustion of poultry-litter-derived hydrochar and biochar. Appl Energy 1080/01431161.2012.682661.
2018;213:510–9. https://doi.org/10.1016/j.apenergy.2017.11.033. [32] Were K, Bui DT, Dick ØB, Singh BR. A comparative assessment of support vector
[8] Tian H, Li J, Yan M, Tong YW, Wang CH, Wang X. Organic waste to biohydrogen: A regression, artificial neural networks, and random forests for predicting and

9
J. Li, et al. Applied Energy 269 (2020) 115166

mapping soil organic carbon stocks across an Afromontane landscape. Ecol Ind apenergy.2012.06.063.
2015;52:394–403. https://doi.org/10.1016/j.ecolind.2014.12.028. [41] Zhao P, Shen Y, Ge S, Chen Z, Yoshikawa K. Clean solid biofuel production from
[33] BenTaieb A, Hamarneh G. Uncertainty driven multi-loss fully convolutional net- high moisture content waste biomass employing hydrothermal treatment. Appl
works for histopathology. Intravasc. Imaging Comput. Assist. Stenting, Large-Scale Energy 2014;131:345–67. https://doi.org/10.1016/j.apenergy.2014.06.038.
Annot. Biomed. Data Expert Label Synth. Springer; 2017. p. 155–63. [42] Danso-Boateng E, Shama G, Wheatley AD, Martin SJ, Holdich RG. Hydrothermal
[34] Vega García M, Aznarte JL. Shapley additive explanations for NO2 forecasting. Ecol carbonisation of sewage sludge: Effect of process conditions on product character-
Inform 2020:56. https://doi.org/10.1016/j.ecoinf.2019.101039. istics and methane production. Bioresour Technol 2015;177:318–27. https://doi.
[35] Lundberg SM, Allen PG, Lee S-I. A Unified Approach to Interpreting Model org/10.1016/j.biortech.2014.11.096.
Predictions; 2017. [43] Zhang L, Wang Q, Wang B, Yang G, Lucia LA, Chen J. Hydrothermal carbonization
[36] Roth A. The Shapley value. Cambridge University Press; 1988. https://doi.org/10. of corncob residues for hydrochar production. Energy Fuels 2015;29:872–6.
1017/cbo9780511528446. https://doi.org/10.1021/ef502462p.
[37] Hart S. Shapley Value. Game Theory, Palgrave Macmillan UK; 1989, p. 210–6. [44] Tripathi M, Sahu JN, Ganesan P. Effect of process parameters on production of
https://doi.org/10.1007/978-1-349-20181-5_25. biochar from biomass waste through pyrolysis: a review. Renew Sustain Energy Rev
[38] Ribeiro MT, Singh S, Guestrin C. “Why should i trust you?” Explaining the pre- 2016;55:467–81. https://doi.org/10.1016/j.rser.2015.10.122.
dictions of any classifier. In: Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data [45] Gascó G, Paz-Ferreiro J, Álvarez ML, Saa A, Méndez A. Biochars and hydrochars
Min., vol. 13–17- August-2016. Association for Computing Machinery; 2016, p. prepared by pyrolysis and hydrothermal carbonisation of pig manure. Waste Manag
1135–44. https://doi.org/10.1145/2939672.2939778. 2018;79:395–403. https://doi.org/10.1016/j.wasman.2018.08.015.
[39] Zhou S, Liang H, Han L, Huang G, Yang Z. The influence of manure feedstock, slow [46] Paneque M, De la Rosa JM, Kern J, Reza MT, Knicker H. Hydrothermal carboni-
pyrolysis, and hydrothermal temperature on manure thermochemical and com- zation and pyrolysis of sewage sludges: What happen to carbon and nitrogen? J
bustion properties. Waste Manag 2019;88:85–95. https://doi.org/10.1016/j. Anal Appl Pyrol 2017;128:314–23. https://doi.org/10.1016/j.jaap.2017.09.019.
wasman.2019.03.025. [47] Salcedo-Sanz S, Cornejo-Bueno L, Prieto L, Paredes D, García-Herrera R. Feature
[40] Chiang KY, Chien KL, Lu CH. Characterization and comparison of biomass produced selection in machine learning prediction systems for renewable energy applications.
from various sources: Suggestions for selection of pretreatment technologies in Renew Sustain Energy Rev 2018;90:728–41. https://doi.org/10.1016/j.rser.2018.
biomass-to-energy. Appl Energy 2012;100:164–71. https://doi.org/10.1016/j. 04.008.

10

You might also like