Journal of Environmental Treatment Techniques  
2019, Special Issue on Environment, Management and Economy, Pages: 1068-1073  
J. Environ. Treat. Tech.  
ISSN: 2309-1185  
Journal web link: http://www.jett.dormaj.com  
Power Density Analysis by using Soft Computing  
Techniques for Microbial Fuel Cell  
Karan Singh* and Dharmendra  
Department of Civil Engineering, National Institute of Technology, Hamirpur (HP) – 177005 – India  
Abstract  
Microbial electrochemical technologies, i.e. microbial fuel cells (MFCs), are attracting attention due to their dual functions of  
energy generation and waste removal from wastewaters. Microorganisms use microbial metabolism to convert biochemical  
metabolic energy into electricity using different substrates. This study was conducted on single-cell microbial fuel cells using  
domestic wastewater as a substrate. A steady state condition has been observed for performance of microbial fuel cell at  
laboratory scale; during this period the maximum voltage, current and power obtained were 210 mV, 24mA and 34.97 mW  
respectively. The experimental results were analyzed to determine the power density using four soft-computing techniques, i.e.,  
an artificial neural network (ANN), an M5P model tree, a random forest (RF) and the Gaussian process (two kernel puk (Pearson  
universal kernel) and rbf (radial basis function)). 70% of the data was used for training and the remaining 30% for validation. The  
input parameters used were day, time, voltage and current, while power density was an output parameter for all modeling  
approaches. Better results were achieved with the artificial neural network, random forest and M5P model tree than the Gaussian  
2
process. The artificial neural network technique gave good results with correlation coefficient (R ), root-relative square error  
(RRSE), relative absolute error (RAE), and root mean square error (RMSE) of 1, 0.4482, 0.4377 and 0.0056, respectively.  
Keywords: Artificial neural network; Gaussian process; M5P model tree; Microbial fuel cell; power density; Random forest  
1
compartment acts as an electron acceptor with electrons  
and hydrogen ions to forms water (9).  
1
Introduction  
The energy requirements of developing countries like  
Microbial growth and the efficiency of reactions within  
the cell are greatly affected by the material used. Different  
studies have used carbon-based materials such as carbon  
fiber and carbon nanotube-based composites (10-12),  
carbon paper (13), carbon felt (14-15), stainless steel mesh  
India, Pakistan, etc., for the domestic and industrial sectors,  
increases almost daily. The International Energy Agency  
(
IEA) has predicted that by 2035 world’s power  
requirements are likely to have risen to 18,000 million  
metric tonnes of oil equivalent (MTOE) from the current  
(
(
16) and carbon cloth (17), as the MFC anode and cathode  
18-21).  
1
2,000 (1). The growing energy requirements from non-  
renewable sources have led to the need to find new,  
renewable and cost-effective resources (2). Fuel cells are  
renewable as well as a green option for electricity  
generation (3). Recently, researchers have used bacteria-  
based microbial fuel cells (MFCs) to convert organic matter  
to electrical energy (4-5). Wastewater acts as a substrate for  
microbial growth (Daniel et al., 2009).  
Potter (6) was the first to generate electrical energy via  
microbes that oxidize organic matter in wastewater (7).  
MFCs help solves major concerns in wastewater treatment  
and energy production. The microbes present in the  
substrate degrade the organic matters and produce energy,  
simultaneously, purifies the wastewater (8). MFCs have  
two compartments, anode and cathode, separated by a  
membrane or a salt bridge. A biofilm develops in the anode  
compartment and acts as a catalyst, transforming the  
biochemical energy from the organic matter to hydrogen  
ions and electrons. The oxygen present in the cathode  
Some researchers have successfully applied, various  
soft computing techniques like Gaussian process, support  
vector regression, M5P model tree, Random forest, and  
artificial neural network, in environmental engineering  
applications (22-24). However, the capacity of these  
techniques in predicting the power density of MFCs has not  
been investigated. In this study, a single chamber microbial  
fuel was fabricated for energy generation. Energy  
generating parameters were obtained i.e., voltage, current,  
and power. After the experiment work, four techniques –  
M5P, ANN, RF, and GP with rbf or puk kernel – were  
developed and used to predict power density of an MFC.  
All for soft computing techniques were compared for best  
suitable technique.  
2 MFC operation  
A single chamber microbial fuel cell (SCMFC) was  
used in this study. Its compartments comprised a 6.5 cm  
diameter acrylic pipe.  
A Nafion-117, PEM (proton  
exchange membrane) separating membrane (Sigma  
Aldrich) was used – i.e., a carbon cloth was used as the  
electrode material for both cathode and anode.  
Corresponding author: Karan Singh, Department of Civil  
Engineering, National Institute of Technology, Hamirpur  
(HP) – 177005 – India. E-mail: karans72@gmail.com.  
1
068  
Journal of Environmental Treatment Techniques  
2019, Special Issue on Environment, Management and Economy, Pages: 1068-1073  
where, P = power, I = current; PD = power density, E =  
voltage, and A = electrode surface area. The voltage and  
current were measured for 22 days after which the fuel cell  
was exhausted. The pH of the substrate feed was  
maintained at between 6.5 and 8 using di-potassium  
hydrogen sulfate. The current and voltage continuously  
increase up-to first 5 days due to decomposition of organic  
matter in the anodic chamber. The maximum obtained  
values for current, voltage and power were 24 mA, 210 mV  
and 34.97mW respectively.  
2
.1 Dataset  
Table 1 gives details of the training and testing  
datasets. Some 102 observations were made, of which 70%  
were used for training and 30% for validation. Input data  
comprised the day and time, and the voltage (V) and  
current (I). The output was the power density (P ).  
D
3
Predictive Modeling Techniques  
3
.1 M5P Model Tree  
M5P model tree uses linear regression, where  
continuous numerical attributes are anticipated. The model  
tree is generated in two stages, i.e., splitting then removal  
of over fitting. Initially, the decision tree is generated by  
the split criteria method, wherein the SD of the class value  
is treated (25). The child node has a smaller standard  
deviation than the parent because of the splitting criteria,  
which ensure that the errors are minimized. Over fitting  
may rises from data splitting due to excess tree growth and  
pruning are used in the final stages to deal with this. Excess  
tree growth is reduced by interchanging the sub-trees using  
linear regression (LR) function – see Equation 3:  
Figure 1: Constructed single-chamber MFC  
The cathode was left uncovered and placed after the  
PEM. The anode was kept 2 cm from the cathode and the  
anodic chamber’s working volume was taken to be 180 ml.  
The circuit was completed with copper wire. Throughout  
incubation, the chamber was kept airtight. The substrate  
was anaerobic digester sludge obtained from biogas  
treatment plant. The substrate was fed with 10ml on  
alternate days. The Chemical oxygen demand of the  
substrate was 34800mg/l initially. The cell voltage (V) and  
current (I) were recorded four times daily using a digital  
multimeter (HTC DM-97, India). Using the current and  
voltage values, the power and power density were  
calculated using equations 1 and 2, respectively:  
|
ꢓꢔ|  
ꢆꢇꢇꢈꢇ ꢇꢉꢊꢋꢌꢍꢎꢈꢏ ꢁ ꢐꢊ(ꢑ) ꢒ ∑  
ꢐꢊ (ꢑ )  
(3)  
ꢕ
|
ꢓ|  
where K illustrates a set of samples that reach the node, sd  
th  
is the standard deviation, and K the i result of a subset of  
i
samples of the potential set.  
3
.2 Artificial Neural Network (ANN)  
ANN works on the concept of brain architecture and  
nervous system functioning. It consists of input and hidden  
layers, iteration, and an output layer. Every layer is  
subdivided into several nodes and links between nodes are  
represented by weighted connections between the layers.  
ꢀ
ꢁ ꢂꢃ  
(1)  
(2)  
ꢄ
P ꢁ  
D
ꢅ
Table 1: Model input parameters  
Training Data  
Validation Data  
Input parameter  
Min.  
1
Max.  
21  
Mean  
11.90  
3.41  
SD  
Min.  
1
Max.  
Mean  
9.22  
SD  
Day  
7.19  
1.76  
30.10  
0.005  
20  
5.87  
1.59  
27.64  
0.004  
Time (Hours)  
1
8
1
7
2.96  
V
I
97  
207  
0.026  
154.21  
0.016  
88  
209  
0.027  
146.32  
0.015  
0.008  
0.007  
1
069  
Journal of Environmental Treatment Techniques  
2019, Special Issue on Environment, Management and Economy, Pages: 1068-1073  
The nodes distribute the input data to the network and  
input layers do not interrupt the overall processing. Data  
processing is a function of the hidden layer and the final  
processing unit is the output layer. The input layer passes  
through the nodes and multiplied by an equivalent weight  
and net output is as following (26) – see Equation 4.  
3.4 Gaussian Process Regression (GPR)  
The primary objective of GPR is to develop systems  
that can predict variables on the basis of past experience.  
The technique is used widely in medicine, engineering, the  
physical sciences, chemistry, etc. The Gaussian process is  
based on the probability which makes predictions of input  
data and gives accuracy for probable variances. The  
statistical significance of the prediction is elevated  
extensively by these estimated variances. GP can extend in  
vast dimensionality. GP is followed by multivariate  
distribution system which produces data by using random  
domain of subset ranges.  
ꢖ = ∑ ꢘ ꢛ ꢜꢙ  
(4)  
ꢗ
ꢙ
ꢙꢚ  
where, Mab = weight of inter-connection from unit a to b, ya  
input value, and zb = outcomes through activation  
function to generate an output for unit b.  
=
3
.3 Random Forest (RF)  
A well-arranged collection of tree predictors generated  
4 Results and discussion  
2
The correlation coefficient (R ), root-relative square  
from input vectors using random vector samples is termed  
an RF approach. Different variables are arranged at each  
node to develop a tree using arbitrarily chosen input  
parameters. A training data set is drawn from randomly  
selected parameters for constructing particular trees and a  
Gini index used to measure the impurity of parameters  
compared to the output (27). RF regression requires two  
pre-defined user variables, an input parameter (m) to be  
used at a separate node to produce a tree and the other the  
number of trees grown (k) (28). The approach is based on  
hit and trial, the variables being selected through the best  
split.  
error (RRSE), relative absolute error (RAE) and root mean  
square error (RMSE) values were used in the study’s  
modeling trails. In the ANN approach, four parameters –  
learning rate, momentum, iteration, and hidden layer –  
must be determined for a given dataset. Several trials  
showed that values of 0.2, 0.1, 4000 and 4 respectively  
were the best combination. The optimal values for user-  
defined parameters were also identified by trial and error  
for the other techniques. Table 2 is a summary of the  
optimal values for user-defined parameters for the different  
algorithms tested.  
Table 2: Ideal input variable values for soft computing  
Variable values  
Data mining technique  
GP (puk kernel)  
GP (rbf kernel)  
M5P  
Gaussian noise = 0,  
Gaussian noise = 0,  
M = 4  
= 1,  
= 1  
= 0.01  
ANN  
Learning rate = 0.2, momentum = 0.1, iteration = 4000, hidden layer = 4,  
K = 0, m = 2, I = 100  
RF  
Table 3: Performance evolution of training and validation datasets used in the soft computing techniques  
Testing dataset  
Techniques  
2
R
RMSE  
0.0723  
0.0056  
0.1512  
0.3059  
1.0555  
RAE  
5.0795  
0.4377  
9.7668  
20.798  
83.8541  
RRSE  
5.8033  
0.4482  
12.1267  
24.545  
84.6788  
M5P  
ANN  
0.9984  
1
RF  
0.9925  
0.989  
0.9548  
GP (puk kernel)  
GP (rbf kernel)  
Training dataset  
M5P  
ANN  
0.9969  
1
0.0839  
0.0063  
0.0885  
0.2196  
0.9165  
6.4983  
0.5286  
6.0213  
15.7615  
85.1817  
7.8771  
0.5949  
8.3133  
20.6262  
86.0658  
RF  
0.9971  
0.9897  
0.9254  
GP (puk kernel)  
GP (rbf kernel)  
1
070  
Journal of Environmental Treatment Techniques  
2019, Special Issue on Environment, Management and Economy, Pages: 1068-1073  
2
Table 3 shows the values of R , RRSE, RAE, and  
from ANN. A similar comparison – Figure 4 – of the  
different modeling techniques suggests that ANN methods  
perform well and can be used for the prediction of power  
density.  
RMSE with training and validation datasets. Higher values  
of R and lower values of RRSE, RAE & RMSE suggest  
that all models, except GP with rbf or puk kernels, work  
well with this dataset. ANN proved to be very accurate and  
can be used effectively to predict power density in an MFC.  
Table 3 shows that ANN had the highest value of R , i.e. 1,  
and the lowest RRSE, RAE and RMSE – 0.4482, 0.4377 &  
.0056, and 0.5949, 0.5286 & 0.0063 – in the training and  
2
4.1 Sensitivity Analysis  
2
Sensitivity analysis (SA) was carried out with the ANN  
model to determine the relative significance of each input  
parameter in predicting MFC power density. Many factors  
affect the magnitude of power density like day, time, V and  
I but V and I were the most important. The SA of ANN was  
shown in Table 4 which considered by eliminating one and  
two parameters in each case, and observing the impact on  
0
validation datasets respectively.  
Figures 2 and 3 show graphs of the predicted versus the  
observed values of MFC power density for the different  
models. (The + 10 % error lines were also shown.) As can  
be seen, the points for M5P, RF and the two GP regressions  
all tend to plot further from the agreement line than those  
2
power density prediction by estimating it in terms of R ,  
RMSE, RAE, and RRSE.  
ANN  
M5P  
RF  
GP_puk  
GP_rbf  
5
4
3
2
1
0
0
1
2
3
4
5
Actual power density (mW/m2)  
Figure 2: Observed vs predicted values of power density, using data mining with the training dataset  
ANN  
M5P  
RF  
GP_puk  
GP_rbf  
5
4
3
2
1
0
0
1
2
3
4
5
Actual power density (mW/m2)  
Figure 3: Observed vs predicted values of power density, using data mining with the validation dataset  
1
071  
Journal of Environmental Treatment Techniques  
2019, Special Issue on Environment, Management and Economy, Pages: 1068-1073  
Actual  
ANN  
M5P  
RF  
GP_puk  
GP_rbf  
6
Training dataset  
Testing dataset  
5
4
3
2
1
0
0
10  
20  
30  
40  
50  
60  
70  
80  
90  
100  
Data Set  
Figure 4: Comparison of predicted vs actual values  
Table 4: Sensitivity analysis of ANN  
ANN  
Input  
combination  
Input variable  
Removed  
2
R
RMSE  
0.0056  
0.0111  
0.0043  
0.7326  
0.4334  
0.0333  
0.287  
RAE  
RRSE  
0.4482  
0.887  
Day, Time, V, I  
Time, V, I  
Day, V, I  
Day, Time, I  
Day, Time, V  
V, I  
1
0.4377  
0.7772  
0.3218  
42.7189  
33.7832  
2.7119  
23.9979  
56.304  
38.1306  
Day  
Time  
V
1
1
0.3427  
58.7751  
34.775  
2.6736  
23.0219  
57.6239  
35.6981  
0.9208  
0.9343  
0.9997  
0.9738  
0.8294  
0.9368  
I
Day, Time  
Time, V  
V, I  
Day, I  
Day, Time  
Time, V  
0.7182  
0.445  
Day, I  
5
Conclusions  
Acknowledgment  
This study was supported by the Government of India,  
Ministry of Human Resource Development. Authors thank  
the colleagues from the National Institute of Technology,  
NIT Hamirpur (H.P.) who provided insight and expertise  
that greatly assisted the current study.  
The accurate prediction of power density in any MFC  
directly affects the substrate treatment efficiency and  
energy generation. In this study, ANN, M5P, RF, and GP  
with either rbf or puk kernel, were all used to predict MFC  
power density using the same four input variables. The  
results show that all models, except GP with either kernel,  
can predict MFC power density most accurately. The ANN  
Ethical issue  
Authors are aware of, and comply with, best practice in  
publication ethics specifically with regard to authorship  
2
model works better than the others (R = 1, RRSE =  
0
.5949, RAE = 0.5286 and RMSE = 0.0063). The ANN  
approach can be applied to all datasets from different  
regions before power density prediction. Sensitivity  
analysis suggests that voltage (V) and current (I) was the  
important parameters for predicting MFC power density.  
The artificial neural network predicts better results from all  
other soft computing techniques, thus for prediction of  
power density ANN soft computing technique can be  
suggested.  
(avoidance of guest authorship), dual submission,  
manipulation of figures, competing interests and  
compliance with policies on research ethics. Authors  
adhere to publication requirements that submitted work is  
original and has not been published elsewhere in any  
language.  
1
072  
Journal of Environmental Treatment Techniques  
2019, Special Issue on Environment, Management and Economy, Pages: 1068-1073  
1
1
1
5. Deng Q, Li X, Zuo J, Ling A, Logan BE (2010) Power  
generation using an activated carbon fiber felt cathode in an  
upflow microbial fuel cell. Journal of Power Sources, 195(4),  
Conflict of Interest  
The author declares that there is no conflict of interests  
regarding the publication of this manuscript. In addition,  
the ethical issues, including plagiarism, informed consent,  
misconduct, data fabrication and/or falsification, double  
publication and/or submission, and redundancy have been  
completely observed by the authors.  
1
130-1135.  
6. Dumas C, Mollica A, Féron D, Basséguy R, Etcheverry L,  
Bergel A (2007) Marine microbial fuel cell: use of stainless  
steel electrodes as anode and cathode materials.  
Electrochimica Acta, 53(2), 468-473.  
7. Ishii SI, Shimoyama T, Hotta Y, Watanabe  
Characterization of filamentous biofilm community  
established in cellulose-fed microbial fuel cell. BMC  
Microbiology, 8(1), 6.  
K (2008)  
a
Authors’ contribution  
All authors of this study have a complete contribution  
for data collection, data analyses and manuscript writing  
a
18. Das S, Ghangrekar MM (2018) Value added product recovery  
and carbon dioxide sequestration from biogas using microbial  
electrosynthesis.  
References  
1
2
2
9. Das S, Chatterjee P, Ghangrekar MM (2018) Increasing  
methane content in biogas and simultaneous value added  
product recovery using microbial electrosynthesis. Water  
Science and Technology, 77(5), 1293-1302.  
1
.
Chu S, Majumdar A (2012) Opportunities and challenges for a  
sustainable energy future. Nature, 488(7411), 294.  
2
.
Kumar A, Kumar N, Baredar P, Shukla A (2015) A review on  
biomass energy resources, potential, conversion and policy in  
India. Renewable Sustainable Energy Rev., 45: 530-539.  
Rahimnejad M, Ghoreyshi AA, Najafpour G, Jafary T (2011)  
Power generation from organic substrate in batch and  
continuous flow microbial fuel cell operations. Applied  
Energy, 88(11), 3999-4004.  
0. Bhowmick GD, Das S, Adhikary K, Ghangrekar MM, Mitra A  
(2019) Using rhodium as a cathode catalyst for enhancing  
3
4
5
6
.
.
.
.
performance of microbial fuel cell. International Journal of  
Hydrogen Energy, 44(39), 22218-22222.  
1. Das S, Ghangrekar MM (2019) Tungsten oxide as  
electrocatalyst for improved power generation and wastewater  
treatment in microbial fuel cell. Environmental technology, 1-  
Angenent LT, Karim K, Al-Dahhan MH, Wrenn BA,  
Domíguez-Espinosa R (2004) Production of bioenergy and  
8
.
biochemicals  
from  
industrial  
and  
agricultural  
2
2
2. Sihag P, Tiwari NK, Ranjan S (2017) Modelling of infiltration  
of sandy soil using gaussian process regression. Modeling  
Earth Systems and Environment, 3(3), 1091-1100.  
3. Saghebian SM, Sattari MT, Mirabbasi R, Pal M (2014)  
Ground water quality classification by decision tree method in  
Ardebil region, Iran. Arabian Journal of Geosciences, 7(11),  
wastewater. TRENDS in Biotechnology, 22(9), 477-485.  
Liu H, Logan BE (2004) Electricity generation using an air-  
cathode single chamber microbial fuel cell in the presence and  
absence of  
a proton exchange membrane. Environmental  
Science & Technology, 38(14), 4040-4046.  
Potter MC (1911) Electrical effects accompanying the  
decomposition of organic compounds. Proceedings of the  
Royal Society of London. Series B, Containing Papers of a  
Biological Character, 84(571), 260-276.  
4
767-4777.  
2
2
4. Singh B, Sihag P, Singh K (2017) Modelling of impact of  
water quality on infiltration rate of soil by random forest  
regression. Modeling Earth Systems and Environment, 3(3),  
7
8
.
.
Lettinga G, Rebac S, Zeeman G (2001) Challenge of  
psychrophilic anaerobic wastewater treatment. TRENDS in  
Biotechnology, 19(9), 363-370.  
Lu N, Zhou SG, Zhuang L, Zhang JT, Ni JR (2009) Electricity  
generation from starch processing wastewater using microbial  
fuel cell technology. Biochemical Engineering Journal, 43(3),  
9
99-1004.  
5. Quinlan JR (1992) Learning with continuous classes. In 5th  
Australian joint conference on artificial intelligence, Hamburg,  
Germany (Vol. 92, pp. 343-348).  
2
2
6. Haykin  
Foundation. Second Edition, Upper Saddle River, New Jersey.  
7. Breiman L, Friedman J, Olshen R, Stone (1984)  
S (1999) Neural Networks: a Comprehensive  
2
46-251.  
C
9
1
1
.
Zhou M, Wang H, Hassett DJ, Gu T (2013) Recent advances  
in microbial fuel cells (MFCs) and microbial electrolysis cells  
Classification and regression trees–crc press. Boca Raton,  
Florida.  
8. Breiman L (1999) Random forests. UC Berkeley TR567.  
(
MECs) for wastewater treatment, bioenergy and  
bioproducts. Journal of Chemical Technology and  
Biotechnology, 88(4), 508-518.  
2
0. Zhang Y, Sun J, Hu Y, Li S, Xu Q (2012) Bio-cathode  
materials evaluation in microbial fuel cells: a comparison of  
graphite felt, carbon paper and stainless steel mesh  
materials. International Journal of Hydrogen Energy, 37(22),  
1
6935-16942.  
1. Ghasemi M, Shahgaldi S, Ismail M, Kim BH, Yaakob Z, Daud  
WRW (2011) Activated carbon nanofibers as an alternative  
cathode catalyst to platinum in a two-chamber microbial fuel  
cell. International Journal of Hydrogen Energy, 36(21),  
1
3746-13752.  
1
1
1
2. Sharma T, Reddy ALM, Chandra TS, Ramaprabhu S (2008)  
Development of carbon nanotubes and nanofluids based  
microbial fuel cell. International Journal of Hydrogen  
Energy, 33(22), 6749-6754.  
3. Yuan Y, Kim SH (2008) Improved performance of a microbial  
fuel cell with polypyrrole/carbon black composite coated  
carbon paper anodes. Bulletin of the Korean Chemical  
Society, 29(7), 1344-1348.  
4. Lv Z, Xie D, Yue X, Feng C, Wei C (2012) Ruthenium oxide-  
coated carbon felt electrode: a highly active anode for  
microbial fuel cell applications. Journal of Power  
Sources, 210, 26-31.  
1
073