Energy 213 (2020) 118806

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Evaluation of artificial neural network algorithms for predicting the

effect of the urine flow rate on the power performance of microbial
fuel cells
A. de Ramo
n-Ferna
ndez a, M.J. Salar-García b, *, D. Ruiz Ferna

ndez a, J. Greenman b,
b, **
I.A. Ieropoulos
a
Department of Computer Technology, University of Alicante, Alicante, E-03690, Spain
b
Bristol BioEnergy Centre, Bristol Robotic Laboratory, Block T, University of the West of England, Bristol, Coldharbour Lane, Bristol, BS16 1QY, UK

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

Article history: Microbial fuel cell (MFC) power performance strongly depends on the biofilm growth, which in turn is
Received 27 March 2020 affected by the feed flow rate. In this work, an artificial neural network (ANN) approach has been used to
Received in revised form simulate the effect of the flow rate on the power output by ceramic MFCs fed with neat human urine. To
20 August 2020
this aim, three different second-order algorithms were used to train our network and then compared in
Accepted 6 September 2020
Available online 9 September 2020
terms of prediction accuracy and convergence time: Quasi-Newton, Levenberg-Marquardt, and Conju-
gate Gradient. The results showed that the three training algorithms were able to accurately simulate
power production. Amongst all of them, the Levenberg-Marquardt was the one that presented the
Keywords:
Artificial neural networks
highest accuracy (R ¼ 95%) and the fastest convergence (7.8 s). These results show that ANNs are useful
Modelling and reliable tools for predicting energy harvesting from ceramic-MFCs under changeable flow rate
Microbial fuel cells conditions, which will facilitate the practical deployment of this technology.
Urine © 2020 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license
Flow rate (http://creativecommons.org/licenses/by/4.0/).
Bioenergy

1. Introduction properties such as high conductivity and nitrogen content, are

benficial for green energy production [4e6].
Microbial Fuel Cell (MFC) technology relies on bacterial meta- In order to facilitate the scalability of this technology, and
bolism to turn the chemical energy stored in a substrate into consequently its real implementation, much progress has been
electricity. The ability of microbes to degrade organic matter allows made in terms of material development, reactor design and opti-
these systems to produce bioenergy and treat wastes simulta- misation. The search for new materials focuses on reducing the cost
neously [1,2]. Respiring bacteria drive the oxidation of the substrate and improving the energy efficiency of the overall system whereas
in the anodic compartment whereas in the cathode takes place the the novel MFC designs aim to facilitate the transition from the
reduction of an oxidant, usually oxygen. Anodic and cathodic laboratory scale to its practical deployment, reducing the mainte-
compartments are physically separated by a selective membrane, nance requirements [7]. The modification of the anode with
with both electrodes being connected through an external circuit conductive polymers has been demonstrated to be a suitable way to
which allows the electrons flow from the anode to the cathode [3]. enhance the energy harvesting from MFCs. Wang et al. (2020) re-
One of the most important benefits of MFCs over other technologies ported to reach up to 515 mW cm2 by using self-supporting pol-
is the possibility of exploiting a broad variety of substrates of waste yaniline-sodium alginate/carbon brush (PANI-SA/CB) hydrogel as
nature, e.g. domestic and industrial wastewaters. Among them, an anode (1.38 times higher than the bare anode) [8]. The same
human urine has gained much attention as a feedstock for these authors improved these results by integrating self-supporting
types of bioelectrochemical systems. Its abundance and natural polypyrrole-carboxymethyl cellulose-titanium nitride/carbon
brush hydrogel (PPy-CMC-TiN/CB). In this case, the modified bio-
anodes allowed to increase the power output up to 4.72 times,
* Corresponding author. compared with the bare electrode [9]. Low-cost catalysts have also
** Corresponding author. been synthesised as alternative materials to expensive noble metals
E-mail addresses: [email protected] (M.J. Salar-García), ioannis. for accelerating the oxygen reduction reaction on the cathode. For
[email protected] (I.A. Ieropoulos).

https://doi.org/10.1016/j.energy.2020.118806
0360-5442/© 2020 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

ndez, M.J. Salar-García, D. Ruiz Ferna

n-Ferna
A. de Ramo
ndez et al. Energy 213 (2020) 118806

instance, recently Xin et al. (2020) elaborated Cu2O decorated designed 6-machine learning algorithms based on 4-input vari-
reduced graphene oxide composite (Cu2O/rGO) to be used as a ables to identify the feed substrate from genomic data. The model
catalyst in MFCs. The 3D cathodes designed allowed MFCs not only showed a maximum accuracy of 93 ± 6% by using an ANN trained
to reach a higher value of power output than using a platinum- on datasets classified at the phylum taxonomic level [21]. This
based catalyst (up to 1362 mW cm2) but also improve the methodology has also been used to predict the anodic biofilm
wastewater treatment capacity (71.5% of chemical oxygen demand communities and MFC performance caused by changes in the
removal) [10]. The combination of iron and nitrogen has also been feedstock composition with a total of 33-tests performed. In this
used to synthesise alternative catalysts for the oxygen reduction case, the authors were able to predict the power density output by
reaction in MFCs. Gadja et al. (2018) reported that the performance the system with a minimum error of 5.76 ± 3.16% when the data
of MFCs enhaced more than 68% when aminoantipyrine is used as a were taxonomically classified at the family level [22].
nitrogen precursor for preparing FeeN-based catalysts, compared The stabilisation period of the power output by MFCs has also
with those working with cathodes containing activated carbon [11]. been predicted by ANNs. This model reported that the power
However, when streptomycin is used as a nitrogen precursor to generation by these systems needs between 12 and 16 weeks to be
elaborate similar cathodes, the power output by the MFCs stable [22]. One of the most recent work is published by Tsompanas
increased up to 74% compared with those using activated carbon as et al. [23] and involves the study of ceramic MFCs fed with a real
a catalyst [12]. waste stream. In this work, ANNs are used to simulate the polar-
The advancements in new materials and designs not only isation curves of different MFC set-ups. With a total number of 264
improve the power performance of this technology but also pro- experiments, the authors’ model reached an accuracy of 99.66% in
mote its use for practical applications such as lighting or powering the data prediction.
electronic devices [13e16]. However, despite the success of these ANNs have also been used to predict the effect of specific design
field trials, most of the improvements reported so far are mainly parameters such as the anode angle with respect to the inlet flow
made by running laboratory assays, which are usually cost and time direction on the electricity generated by MFCs. The experimental
demanding, and rarely represent real-world conditions. For these results obtained by Jaeel et al. [24] showed that MFCs reached the
reasons, the use of mathematical modelling to simulate and opti- maximum power output (486 mW m2) when the anode is
mise MFC performance has gained much attention in the last few perpendicular to the substrate flow direction at the lowest feed
years. These techniques allow us to cover multiple scenarios flow rate. In this case, a three-layer ANN model was used to predict
simultaneously under more realistic conditions. So far, mathe- the efficiency of the systems in terms of power production. The
matical models have been commonly used to model all the phe- model revealed a good-fitting between the experimental and
nomena which take place in an MFC. However, these kinds of simulated data with a correlation coefficient of R ¼ 99.889%. A
models usually require an in-depth knowledge of the system, three-layer ANN model was also used by Ismail et al. [25] in 2017 to
which is difficult in complex systems such as MFCs [17,18]. Alter- predict the power generation of double-chamber MFCs continu-
native modelling techniques based on artificial intelligence (AI) has ously fed with domestic wastewater enriched with giant reed as a
strongly burst in a broad domain such as health, environmental new energy source. The ANN approach presented in this work re-
sciences or biotechnology. ported a good-fitting between the experimental and predicted data
Artificial neural networks (ANNs) represent a very important (R ¼ 99.93%), which shows the potential application of AI modelling
and constantly evolving field within AI. They are mathematical tools to optimise the performing of complex systems such as MFCs.
models inspired by the functioning of biological neural networks. More recently, Ali et al. [26] performed an experimental assay to
They consist of a set of simple processing units called nodes or optimise the MFCs set-up for maximising the electricity generation.
neurons organised in layers with a high level of connection be- The authors studied the effect of different factors, e.g. types of salt
tween them that allow the input signals to be transmitted through bridges as separators, as well as different salt concentrations,
the different layers until generating output values. The following temperature or the surface area of electrodes, on the voltage
elements are in any model based on a neural network approach generated by MFCs. A four-layer feed-forward ANN approach was
[19]: employed by the authors for simulating the voltage generated by
the systems. The model developed was able to predict the voltage
 Input signals (x1, x2, …xn), which may be external data or pro- production with a correlation between the experimental and
vided by other neurons. simulated data of 99.9%.
 Output signals (y1, y2, …ym), which represent the network output In line with the growing interest of the scientific community in
values. this field, in the present work the power performance of ceramic-
 Weights (w1k, w2k, …wnk), which represent the interaction be- based MFCs continuously fed with neat human urine was simu-
tween two neurons. The weights are modified during the lated by ANNs. Terracotta clay membranes with different porosities
training of the network in order to obtain the expected output and bulk resistances, as well as the urine flow rate were used as
values. input variables. 16-different feed flow rates were applied on the
 Biases (W1, W2, …Wk), which are values associated with each node MFCs including 6-different ceramic membranes in triplicate, with a
and allow the activation function to be shifted to the left or right, total number of 288 tests run. In order to obtain the best model to
to better fit the data. predict the MFC power performance, several topologies of ANN
designs were studied and three training algorithms were compared
ANNs can be classified according to different criteria such as according to their convergence velocities in training and perfor-
their topology (single or multilayer), the type of learning (super- mances in testing.
vised or unsupervised) or the type of connection between layers
(feed-forward or feedback) [19]. 2. Materials and methods
As in other research fields, the use of ANNs to model the per-
formance of MFCs is focusing the attention of many researchers, 2.1. Microbial fuel cell set-up
increasing the number of published papers in the last few years
[20]. ANNs have been recently used to predict the feed substrate Cuboid single-chamber MFCs made of acrylic were used to
from 69 different microbial communities. In this work, the authors perform the experiments. The cathodes consisted of a paste of
2

ndez, M.J. Salar-García, D. Ruiz Ferna

n-Ferna
A. de Ramo
ndez et al. Energy 213 (2020) 118806

Table 1 continuously fed with neat human urine at a flow rate of

Depiction of the 96-conditions set run in triplicate to perform the 288 essays. 0.06 mL min1 and the external loading gradually adjusted. After 3
Group Name Bulk resistance (U) Porosity (%) Feed flow rate (mL.min1) months working under these conditions and once the systems were
1 64 26.6 0.06
completely stable and the biofilm well developed through the
2 70.6 26.1 0.19 anode, the feed flow was slowly increased up to 3.88 mL min1 at a
3 92 27.7 0.27 fixed external loading of 500 U. The effect of 16-different feed flows
4 114.2 25.8 0.38 on the power output by the MFCs was experimentally assessed (see
5 124 27.1 0.49
Table 1) with a total number of 288 tests run. A multichannel
6 497.2 16.8 0.58
0.68 Agilent recorder data logger (LXI 34972A data acquisition/Switch
0.72 unit) was used to continuously monitor the MFC voltage.
0.92
0.98
1.18 2.3. Porosity analysis
1.33
1.54 The total porosity of the terracotta clay membranes was ana-
1.75 lysed by mercury intrusion porosimetry (Poremaster-60 GT,
1.95
Quantachrome Instrument, United Kingdom). This device includes
3.88
dual high-pressure transducers, which allow it to improve the ac-
curacy of the results, as well as two built-in automated low-
pressure ports. This system is able to make intrusion/extrusion
activated carbon and polytetrafluoroethylene (PTFE) as a binder,
analysis from vacuum to 60.00 psi.
pressed over a piece of stainless steel (12.25 cm2). Regarding the
anodes, a piece of carbon veil (20 g m2, PRF composites, Dorset,
UK) coated with activated carbon were placed in the anodic 2.4. Impedance spectroscopy (IES)
compartment and connected externally to the cathode through a
chromium-nickel cable [27]. As a membrane, flat square pieces of The IES technique was used to determine the bulk resistance
terracotta clay (12.25 cm2) were cut and kilned by using different (Rb) of the terracotta clay membranes by using mAutoLab III with a
procedures, which involve different combinations of temperature frequency response analyser FRA2. The measurements were per-
and ramp time. The different kilning methods allowed finely tuning formed in the frequency range of 100 kHz to 10 mHz, at AC
membrane properties such as porosity, bulk resistance or pore size amplitude of 10 mV, in a two-electrode configuration. The value of
(see Table 1). The final thickness of the membranes was 3 mm. In Rb was determined by the intersection of the semicircle with the Z’
this work, 6-different ceramic membranes were elaborated and axis in the Nyquist plot [28,29].
assessed as MFC separators in triplicate, as can be seen in Fig. 1.
2.5. Artificial neural network. Multilayer perceptron

2.2. Microbial Fuel Cell inoculation A standard type of ANN, namely multilayer perceptron (MLP),
was selected for the prediction of the power performance by MFCs
The air-breathing single-chamber MFCs were inoculated with a from three input variables: feed flow, bulk resistance, and porosity
solution containing urine and sludge (1:1 v/v) in batch mode, of the membrane (see Fig. 2). The MLP is one of the most used
which was replenished daily during 4 days. Then, the systems were neural network models since it is capable of acting as a universal

Fig. 1. Image of the experimental set-up involving 6-MFC groups run in triplicate and fed at 16 different feed flows with a total number of 288 assays.

3

ndez, M.J. Salar-García, D. Ruiz Ferna

n-Ferna
A. de Ramo
ndez et al. Energy 213 (2020) 118806

by the network for pattern n defined as:

X
M
en ¼ ðsi  yi Þ2 (5)
i¼1

where M is the total number of output nodes, yi the output value

provided by the network at the ith output node and si the observed
or target value at the ith output node. The function error depends
exclusively on the parameters of the network (weights and biases).
Depending on the value that these parameters take, more or less
weight will be given to neurons and their connections with the
successive layers. Therefore, by modifying weights of the input
layer, the influence of each input variable on the output variable
Fig. 2. MLP architecture based on the input and output variables of our system.
will be modified. These parameters can be grouped into a single
vector of the corresponding dimension, which we will denote by w.
approximating function. MLP emerges as an evolution of the single- In this sense, we can write the following expression in order to
layer perceptron developed by Rosenblatt in 1958 [30], which indicate that the value of the error made by the neural network will
allowed solving only linearly separable problems. In 1969, Minsky depend on this vector:
and Papert [31] suggested the combination of several single-layer
en ¼ f ðwÞ (6)
perceptrons to overcome this limitation and solve some non-
linear challenges, thus creating a multilayer neural network of With this formalisation, the aim is to find the value w* for which
forwarding propagation. The neurons in the input layer are a global minimum of the function f is obtained, turning the learning
responsible for receiving signals from outside and propagating problem into an optimisation problem. If the function f meets
these signals to the output layer through the hidden layers. The sufficient derivability conditions, the first and second derivative of
neurons of this layer perform non-linear processing of the received this function provides the gradient vector and the Hessian matrix
signals. Our MLP will consist of a single hidden layer since usually respectively:
most of the existing problems can be solved with this configuration,
which also reduces the computation time. vf
; for i ¼ 1; …; N (7)
The propagation rule determines the potential resulting from vwi
the interaction of neuron i with the n neighbouring neurons. One of
the most common and simple rules is the weighted sum of the
v2 f
inputs with their corresponding weights and biases: Hij f ðwÞ ¼ ; for i; j ¼ 1; …; N (8)
vwi vwj
X
n
neti ¼ wij xj þ qi (1) From these equations, a point (network parameters) can be
j¼1 found for which the gradient vector is null and that represents a
maximum or a minimum of the function f . On the other hand, for a
The activation function f is responsible for sending the value minimum, some conditions of the Hessian matrix must be verified.
obtained in the propagation rule and defining the output of the The most common neural network training methods are based
layer. Here, the tangent sigmoid function was used for the hidden on conditions of derivability of the error function. The gradient
and output layer. It is one of the most used activation functions and descent or backpropagation method [32] is one of the most popular
is defined in the interval [-1, 1] by the following equation: and implemented for its ease and scope of application. This method
propagates the error measured in the output layer to the hidden
2
ftansig ðxÞ ¼ 1 (2) layers, adjusting the network parameters in the steepest descent
1 þ e2x direction, that is, the most negative of the gradients. It is, therefore,
From the equations so far raised, the value of the network a first-order method since it only uses the gradient vector. How-
output can be generalised as follows: ever, although it is a simple algorithm, multiple iterations are
0 1 necessary to verify that the direction chosen is the one that allows
X
n to reduce the error function and achieve the convergence as quickly
yi ¼ f @ wij xj þ qi A (3) as possible.
j¼1 Second-order methods make use of the Hessian matrix to
reduce the number of iterations necessary until reaching the
convergence. However, the main disadvantage of these methods is
the high computational cost involved in the calculation of the
Hessian matrix. To overcome this limitation, alternative methods
2.5.1. Training algorithms have emerged that make modifications in obtaining training di-
Since the objective of the network is to provide an output value rections or that propose approximations of the Hessian matrix to
as close to the observed value as possible, network learning is achieve faster and simpler convergence.
posed as a problem of minimising the output overall error defined Here, some of the most implemented second-order training
by the function: methods were used to test which of them produces better results
and faster training for the case of study. Subsequently, these
1 XN
training algorithms are explained in more detail. To do this, we will
E¼ en (4)
N n¼1 denote as wi the network parameter vector in the iteration i,
fi ¼ f ðwi ) the error value in the iteration i and gi ¼ Vf ðwi Þ the
where N is the number of training patterns and en is the error made gradient value of the error function in the iteration i.
4

ndez, M.J. Salar-García, D. Ruiz Ferna

n-Ferna
A. de Ramo
ndez et al. Energy 213 (2020) 118806

2.5.1.1. Quasi-Newton methods. Newton’s method [33] makes use the fastest convergence occurs. To that end, the CG method uses
of the Hessian matrix to find the best directions of variation of the conjugate training directions that generally produce faster
network parameters. This method allows generating the vector of convergence directions [38]. The set of training directions is
parameters such as: defined by the following equation:

wiþ1 ¼ wi  Hi1 gi (9) diþ1 ¼ giþ1 þ di gi (15)

If the Hessian is not defined positive, the succession of param- where g is the conjugate parameter, for which different authors
eters might tend to a maximum instead of a minimum. To avoid this have proposed alternative methods of calculation [39,40]. Finally,
problem, the above equation may be modified as follows: the parameters of the error function are obtained as follows:
 
wiþ1 ¼ wi  Hi1 gi v (10) wiþ1 ¼ wi þ di v (16)
In the next section, the three training algorithms are compared
Thus, the method attempts to determine the training direction
according to their convergence velocities in training and perfor-
first and then, an appropriate training speed v. However, as
mances in testing.
explained above, the calculation of the Hessian matrix is complex
and implies a high computational cost. The Quasi-Newton (QN)
methods seek an approximation of the inverse of the Hessian ma-
3. Results and discussion
trix without the need to solve the second derivatives in each iter-
ation, thus simplifying its calculation. If we denote this
3.1. Algorithm analysis
approximation for the iteration i as Gi , the Quasi-Newton method
can be defined as follows:
Three MLPs were developed for testing each training algorithm
wiþ1 ¼ wi  ðGi gi Þv (11) and compared its performance by using MATLAB version R2017b as
the computational platform. The optimal number of neurons in the
The most popular methods for the calculation of Gi are the hidden layer cannot be known in advance and must be obtained
Davidon-Fletcher-Powell formula [34,35] and the Broyden- empirically. This is a key factor since it significantly influences the
Fletcher-Goldfard-Shanno formula [36], which will be used to ability of the network to generalise from training data [41]. A small
train our network. number of neurons could lead to poor training (under-fitting) while
too many neurons can lead to overtraining (over-fitting) [42]. That
2.5.1.2. Levenberg-Marquardt algorithm. Like the Quasi-Newton is the reason why different MLP topologies were defined by varying
methods, the Levenberg-Marquart (LM) algorithm [37] was also the number of neurons in the hidden layer from 3 to 12. Each al-
developed to perform faster training without calculating the exact gorithm was tested with each configuration and a total number of
Hessian matrix. When the error function approaches a sum of 100 runs were performed per algorithm and configuration to find
squares as happen in feed-forward networks, then the Hessian the winning model. After each iteration, the values of the weights
matrix can be approximated as follows: were saved and the model was validated through performance
metrics. The experimental data set was divided into 3 subsets: 60%
H ¼ JT J (12) of the data was used for network training with the selected algo-
rithm, 20% to validate the model and warn of under-fitting and
where J is the Jacobian matrix that contains the first derivatives of over-fitting problems, and the remaining 20% to evaluate the ac-
the error function concerning weights and biases. The gradient is curacy of the model. Finally, a maximum number of 100 epochs was
obtained as follows: established and the training was stopped after six consecutive in-
creases in validation error and the best performance was taken
g ¼ JT e (13) from the epoch with the lowest validation error.
Table 2 shows the best results obtained for each training and
where e is the network error vector. The LM method updates the configuration algorithm: the correlation coefficient (R) measures
parameters vector for each iteration by using the following the degree of association between two variables, in this case, the
equation: power value observed and predicted by the network; the mean
squared error (MSE) allows us to know the amount of error that
h i1
wiþ1 ¼ wi  JiT Ji þ mI JiT ei (14) exists between the two data sets of both variables; and CPU time
measures the speed of convergence, which means the time needed
by each algorithm to find a solution that meets the training stop
where m is the learning rate. When m is null, the algorithm becomes
criteria. The topologies that exhibited the best results for the QN,
Newton’s method (Eq. (9)). The LM algorithm seems to be the
LM and CG algorithms were those that had 9, 8 and 11 neurons in
fastest method to train moderately sized feed-forward networks
the hidden layer, respectively.
(up to several hundred parameters). Its main disadvantage is that it
Fig. 3 shows the scatter plot of the winning models of each al-
requires storing the Jacobian matrices that can be very large for
gorithm. For each of them, the fit between the predicted model and
certain data sets. This fact results in a large use of memory.
the observed data was fairly precise, and therefore represented a
strong linear correlation between both data sets. The LM algorithm
2.6. Conjugate Gradient was the one that obtained the highest accuracy in the adjustment
(greater than 95%), with an MSE of 7.90 in the testing period and
The Conjugated Gradient (CG) method seeks to speed up the convergence time of 7.8 s. The QN algorithm obtained an R and MSE
convergence concerning the traditional CG method without of 0.932 and 7.89, respectively, for a time of 11.2 s. On the other
calculating the Hessian matrix. The traditional gradient descent hand, the CG algorithm was the one that obtained the lowest R
method defines as training direction the one that faster minimises (0.92) with an MSE of 10.7 and convergence time of 18.2 s.
the error function. However, this does not necessarily imply that In general, the three models offered good fit accuracy and a
5

ndez, M.J. Salar-García, D. Ruiz Ferna

n-Ferna
A. de Ramo
ndez et al. Energy 213 (2020) 118806

Table 2
Performance results of the training algorithms in the testing period.

Neurons QN LM CG
Hidden layer
R MSE (mW.cm2) Time (s) R MSE (mW. cm2) Time (s) R MSE (mW cm2) Time (s)

3 0.886 10.32 11.4 0.862 9.38 7.2 0.791 9.54 23.2

4 0.888 8.88 10.8 0.861 10.13 9.5 0.869 7.88 16.8
5 0.884 7.56 9.1 0.892 7.94 11.3 0.888 9.51 9.2
6 0.901 10.20 18.0 0.878 7.46 17.3 0.898 8.07 23.6
7 0.894 11.25 13.3 0.928 9.15 14.6 0.856 10.64 27.4
8 0.895 8.45 14.3 0.950 7.90 7.8 0.898 8.56 14.1
9 0.932 7.89 11.2 0.920 8.79 9.6 0.901 8.18 34.3
10 0.779 12.56 8.4 0.946 7.67 13.5 0.913 7.66 27.0
11 0.898 8.65 27.6 0.916 7.80 18.5 0.920 10.07 18.2
12 0.895 7.78 16.6 0.884 7.72 15.4 0.841 10.13 17.6

Fig. 3. Regression plot of observed and predicted power in testing period: (A) QN, (B) LM, (B) CG.

small MSE, indicating little deviation between the observed value 3.2. Analysis of the experimental and simulated results
and the one predicted by the network. In fact, the LM, QN and CG
algorithms made a total of 85%, 82%, and 89% predictions respec- According to the previous discussion, the algorithm which al-
tively, with a relative error lower than 10%. Moreover, the validation lows us to predict the power performance of the MFCs more
error for each algorithm was low and slightly higher than the accurately is LM (R ¼ 0.95%). Fig. 5 shows the experimental results
training error, which rules out under-fitting and overfitting prob- obtained by feeding the 6-groups of different ceramic MFCs with
lems (see Fig. 4). In light of these results, it can be concluded that neat human urine at 16 target feed flow rates. As can be observed,
the LM algorithm was the one that best and fastest solved the po- the experimental power output by the groups 1, 2, 3 and 4 is very
wer prediction problem in MFCs. similar, being group 1, which achieved the maximum value. This

6

ndez, M.J. Salar-García, D. Ruiz Ferna

n-Ferna
A. de Ramo
ndez et al. Energy 213 (2020) 118806

Fig. 4. MSE for each winning model in validation period: (A) QN, (B) LM, (C) CG.

result might be related to the low bulk resistance of the ceramic changes in the feed flow rate. For feeding flow rates ranging be-
membrane of this group (64 U). However, the performance of these tween 0.06 and 0.72 mL min1 the power performance increased
four groups showed a similar trend, probably due to the porosities significantly, showing higher slope that in the rest of groups.
of the ceramic membranes being quite similar and the bulk re- However, once the flow rate was higher than 0.72 mL min1, power
sistances being relatively low. All of them reached the maximum output decreased dramatically, almost at a level lower to that
power output at feeding flows ranged between 1.18 and achieved following the previous increase. These results might be
1.53 mL min1, although MFCs belonging to group 1 were the only related to the blockage of the pores, which could happen easily
which kept increasing the power performance for flow rates higher than in other membranes with higher porosities. In this case, the
than 1.33 mL min1. After these values of flow rate, the power maximum power output was 28.6%, lower than that observed in
performance of these four groups decreased until reaching the group 1 and it was obtained at a feed flow of 0.72 mL min1, less
initial power output or even lower. These results might be due to than the half of flow rate needed by the group 1 to reach the best
low flow rates of fuel supply are limiting and the fluidic shear rate is performance.
low, promoting the formation of thick and dense biofilms, giving
rise to diffusion-limitation of the substrates. By contrast, if the flow 4. Conclusions
rate-shear rate is higher, the thickness of the biofilm is reduced by
the detachment of outer layers of cells, and the current produced The flow rate, and therefore, the hydraulic retention time and
increases due to higher supply rate of limiting nutrient and/or the shear stress significantly affect the power performance of MFCs.
removal of diffusion-limiting outer layers of biofilm. For these For this reason, this work aims at simulating the effect of the neat
reasons, an increase in the flow rate initially facilitates the devel- urine flow rate on the power performance of ceramic MFCs by using
opment of a more efficient biofilm, which improve the power an ANN approach. To this end, a multilayer perceptron, one of the
performance by the system [43]. However, if the flow rate is very most commonly used ANN models worldwide, was designed. Three
high, the current produced might be reduced due to cell detach- second-order algorithms, the QN, LM and CG were used to train our
ment of the inner core layer of cells [44e47]. network. These methods represent an advantage in terms of
In the case of groups 5 and 6, both exhibit lower values of power computational cost regarding other second-order algorithms, such
output than the groups previously commented, probably because as the backpropagation method, since they use approximations of
the bulk resistance of the ceramic membranes used in these cases the Hessian matrix or modifications in the training directions that
are significantly higher. It is worth mentioning that the perfor- speed up the network learning. Each of these algorithms was tested
mance of MFCs working with the least porous membrane (16.8%) for different network topologies until the best model was found. In
and the highest bulk resistance (497.2 U) was more sensitive to general, the three training algorithms tested were able to
7

ndez, M.J. Salar-García, D. Ruiz Ferna

n-Ferna
A. de Ramo
ndez et al. Energy 213 (2020) 118806

Fig. 5. Comparison of the simulated and the mean experimental power performance of ceramic MFCs: A) Group 1, B) Group 2, C) Group 3, D) Group 4, E) Group 5, F) Group 6,
previously described.

accurately simulate the power prediction, showing a small MSE and of this technology.
convergence time, proving to be suitable algorithms for this case
study. Among all of them, the LM was the one that presented the Credit author statement
highest accuracy (R ¼ 95%) and the fastest convergence (7.8 s).
These results show that ANNs are useful and reliable tools for
n-Ferna

ndez and MJ. Salar-García: Conceptualisa-
A. De Ramo
predicting energy harvesting from ceramic-MFCs under changeable
tion, Data curation, Formal analysis, Investigation, Methodology,
flow rate conditions, which will facilitate the practical deployment
Visualisation, Writing - original draft, Writing - review & editing. D.
8

ndez, M.J. Salar-García, D. Ruiz Ferna

n-Ferna
A. de Ramo
ndez et al. Energy 213 (2020) 118806

Ruiz Ferna
ndez: Funding acquisition, Supervision, Writing - re- lights. Appl Energy 2020;277:115514. https://doi.org/10.1016/
j.apenergy.2020.115514.
view & editing. J. Greenman: Supervision, Writing - review & [17] Ortiz-Martínez VM, Salar-García MJ, de los Ríos AP, Hern
andez-Ferna
ndez FJ,
editing I. Ieropoulos: Conceptualisation, Funding acquisition, Egea JA, Lozano LJ. Developments in microbial fuel cell modeling. Chem Eng J
Project administration, Resources, Supervision, Writing - original 2015;271:50e60. https://doi.org/10.1016/j.cej.2015.02.076.
[18] Xia C, Zhang D, Pedrycz W, Zhu Y, Guo Y. Models for microbial fuel cells: a
draft, Writing - review & editing.
critical review. J Power Sources 2018;373:119e31. https://doi.org/10.1016/
J.JPOWSOUR.2017.11.001.
Declaration of competing interest [19] Krose B, Smagt PA Van Der. An introduction to neural networks. 1996.
Amsterdam. Amsterdam: The University of Amsterdam; 1996.
[20] Sewsynker-Sukai Y, Faloye F, Kana EBG. Artificial neural networks: an efficient
The authors declare that they have no known competing tool for modelling and optimization of biofuel production (a mini review).
financial interests or personal relationships that could have Biotechnol Biotechnol Equip 2017;31:221e35. https://doi.org/10.1080/
13102818.2016.1269616.
appeared to influence the work reported in this paper.
[21] Cai W, Lesnik KL, Wade MJ, Heidrich ES, Wang Y, Liu H. Incorporating mi-
crobial community data with machine learning techniques to predict feed
Acknowledgement substrates in microbial fuel cells. Biosens Bioelectron 2019;133:64e71.
https://doi.org/10.1016/j.bios.2019.03.021.
[22] Singh S, Srinivasan K, Chen BY, Singh H, Goyal A, Garg A, et al. A novel method

n Seneca (Ref. 20372/
M.J. Salar-García is supported by Fundacio for determination of a time period for stabilization of power generation of
PD/17) and A. De Ramo
n-Ferna

ndez by the Spanish Ministry of microbial fuel cell with effect of microorganisms. Int J Energy Res 2019;43:
5834e40. https://doi.org/10.1002/er.4685.
Economy and Competitiveness (Ref. BES-2015-073611). I. Ier- [23] Tsompanas MA, You J, Wallis L, Greenman J, Ieropoulos I. Artificial neural
opoulos also thanks the Bill & Melinda Gates Foundation (grant no. network simulating microbial fuel cells with different membrane materials
INV-006499) for its financial support. and electrode configurations. J Power Sources 2019;436:226832. https://
doi.org/10.1016/j.jpowsour.2019.226832.
[24] Jaeel AJ, Al-Wared AI, Ismail ZZ. Prediction of sustainable electricity genera-
References tion in microbial fuel cell by neural network: effect of anode angle with
respect to flow direction. J Electroanal Chem 2016;767:56e62. https://doi.org/
[1] Potter MC. Electrical effects accompanying the decomposition of organic 10.1016/j.jelechem.2016.02.015.
compounds. Proc R Soc B Biol Sci 1911;84:260e76. https://doi.org/10.1098/ [25] Ismail ZZ, Al-wared AI, Jaeel AJ. Recourse recovery of bioenergy from cellulosic
rspb.1911.0073. material in a microbial fuel cell fed with giant reed-loaded wastewater. Bio-
[2] Stirling JL, Bennetto HP, Delaney GM, Mason JR, Roller SD, Tanaka K, et al. fuels 2019;10:737e45. https://doi.org/10.1080/17597269.2017.1409057.
Microbial fuel cells. Biochem Soc Trans 1983;11:451e3. https://doi.org/ [26] Ali, Al-Mussawy HA, Hussein MJ, Hamadi NJ. Experimental and theoretical
10.1042/bst0110451. study on the ability of microbial fuel cell for electricity generation. Pollution
[3] Logan BE, Hamelers B, Rozendal R, Schro €der U, Keller J, Freguia S, et al. Mi- 2018;4:359e68. https://doi.org/10.22059/poll.2017.246092.341.
crobial fuel cells: methodology and technology. Environ Sci Technol 2006;40: [27] Salar-García MJ, Ieropoulos I. Optimisation of the internal structure of ceramic
5181e92. https://doi.org/10.1021/es0605016. membranes for electricity production in urine-fed microbial fuel cells. J Power
[4] Ieropoulos I, Greenman J, Melhuish C. Urine utilisation by microbial fuel cells; Sources 2020;451:227741. https://doi.org/10.1016/j.jpowsour.2020.227741.
Energy fuel for the future. Phys Chem Chem Phys 2012;14:94e8. https:// [28] Bellucci F, Valentino M, Monetta T, Nicodemo L, Kenny J, Nicolais L, et al.
doi.org/10.1039/c1cp23213d. Impedance spectroscopy of reactive polymers. 1. J Polym Sci, Part B: Polym
[5] Gajda I, Obata O, Jose Salar-Garcia M, Greenman J, Ieropoulos IA. Long-term Phys 1994;32:2519e27. https://doi.org/10.1002/polb.1994.090321509.
bio-power of ceramic microbial fuel cells in individual and stacked configu- [29] Salar-García MJ, Ortiz-Martínez VM, de los Ríos AP, Hern
andez-Ferna
ndez FJ.
rations. Bioelectrochemistry 2020;133:107459. https://doi.org/10.1016/ A method based on impedance spectroscopy for predicting the behavior of
j.bioelechem.2020.107459. novel ionic liquid-polymer inclusion membranes in microbial fuel cells. En-
[6] Santoro C, Garcia MJS, Walter XA, You J, Theodosiou P, Gajda I, et al. Urine in ergy 2015;89:648e54. https://doi.org/10.1016/j.energy.2015.05.149.
bioelectrochemical systems: an overall review. ChemElectroChem 2020;7: [30] Rosenblatt F. The perceptron: a probabilistic model for information storage
1312e31. https://doi.org/10.1002/celc.201901995. and organization in the brain. Psychol Rev 1958;65:386e408. https://doi.org/
[7] Santoro C, Arbizzani C, Erable B, Ieropoulos I. Microbial fuel cells: from fun- 10.1037/h0042519.
damentals to applications. A review. J Power Sources 2017;356:225e44. [31] Minsky M, Perceptrons Papert S. An introduction to computational geometry.
https://doi.org/10.1016/J.JPOWSOUR.2017.03.109. Science 1969;165:780e2. https://doi.org/10.1126/science.165.3895.780. 80-.
[8] Wang Y, Wen Q, Chen Y, Zheng H, Wang S. Enhanced performance of mi- [32] Rumelhart DE, Hinton GE, Williams RJ. Learning representations by back-
crobial fuel cell with polyaniline/sodium alginate/carbon brush hydrogel propagating errors. Nature 1986;323:533e6. https://doi.org/10.1038/
bioanode and removal of COD. Energy 2020;202:117780. https://doi.org/ 323533a0.
10.1016/j.energy.2020.117780. [33] Verbeke J, Cools R. The Newton-raphson method. Int J Math Educ Sci Technol
[9] Wang Y, Wen Q, Chen Y, Li W. Conductive polypyrrole-carboxymethyl cellu- 1995;26:177e93. https://doi.org/10.1080/0020739950260202.
lose-titanium nitride/carbon brush hydrogels as bioanodes for enhanced en- [34] Fletcher R, Powell MJD. A rapidly convergent descent method for minimiza-
ergy output in microbial fuel cells. Energy 2020;204:117942. https://doi.org/ tion. Comput J 1963;66:163e8. https://doi.org/10.1093/comjnl/6.2.163.
10.1016/j.energy.2020.117942. [35] Davidon WC. Variable metric method for minimization. SIAM J Optim
[10] Xin S, Shen J, Liu G, Chen Q, Xiao Z, Zhang G, et al. High electricity generation 1991;1e17. https://doi.org/10.1137/0801001.
and COD removal from cattle wastewater in microbial fuel cells with 3D air [36] Fletcher R. Practical methods of optimization. John Wiley & Sons, Inc.; 1986.
cathode employed non-precious Cu2O/reduced graphene oxide as cathode https://doi.org/10.1002/adma.200400767.
catalyst. Energy 2020;196:117123. https://doi.org/10.1016/ [37] More
JJ. The Levenberg-Marquardt algorithm: implementation and theory.
j.energy.2020.117123. 1978. p. 105e16. https://doi.org/10.1007/bfb0067700.
[11] Gajda I, Greenman J, Santoro C, Serov A, Melhuish C, Atanassov P, et al. [38] Adeli H. Machine learning - neural networks, genetic algorithms and fuzzy
Improved power and long term performance of microbial fuel cell with Fe-N- systems. Kybernetes 1999;28:317e8.
C catalyst in air-breathing cathode. Energy 2018;144:1073e9. https://doi.org/ [39] Fletcher R. Function minimization by conjugate gradients. Comput J 1964;7:
10.1016/j.energy.2017.11.135. 149e54. https://doi.org/10.1093/comjnl/7.2.149.
[12] Salar Garcia MJ, Santoro C, Kodali M, Serov A, Artyushkova K, Atanassov P, [40] Polak E, Ribiere G. Note sur la convergence de me
thodes de directions con-
et al. Iron-streptomycin derived catalyst for efficient oxygen reduction reac- jugue
es. Rev Française d’informatique Rech Ope
rationnelle Se
rie Rouge
tion in ceramic microbial fuel cells operating with urine. J Power Sources 1969;3:35e43. https://doi.org/10.1051/m2an/196903r100351.
2019;425:50e9. https://doi.org/10.1016/J.JPOWSOUR.2019.03.052. [41] Kavzoglu T. An investigation of the design and use of feed-forward artificial
[13] Walter XA, Merino-Jime
nez I, Greenman J, Ieropoulos I. PEE POWER® urinal II neural networks in the classification of remotely sensed images. University of
- urinal scale-up with microbial fuel cell scale-down for improved lighting. Nottingham; 2001.
J Power Sources 2018;392:150e8. https://doi.org/10.1016/ [42] Panchal G, Ganatra A, Kosta YP, Panchal D. Behaviour analysis of multilayer
j.jpowsour.2018.02.047. perceptronswith multiple hidden neurons and hidden layers. Int J Comput
[14] Walter XA, Stinchcombe A, Greenman J, Ieropoulos I. Urine transduction to Theory Eng 2011;3:332e7. https://doi.org/10.7763/ijcte.2011.v3.328.
useable energy: a modular MFC approach for smartphone and remote system [43] Juang DF, Yang PC, Kuo TH. Effects of flow rate and chemical oxygen demand
charging. Appl Energy 2017;192:575e81. https://doi.org/10.1016/ removal characteristics on power generation performance of microbial fuel
j.apenergy.2016.06.006. cells. Int J Environ Sci Technol 2012;9:267e80. https://doi.org/10.1007/
[15] Walter XA, Greenman J, Ieropoulos IA. Microbial fuel cells directly powering a s13762-012-0032-z.
microcomputer. J Power Sources 2020;446:227328. https://doi.org/10.1016/ [44] Ieropoulos I, Winfield J, Greenman J. Effects of flow-rate, inoculum and time
j.jpowsour.2019.227328. on the internal resistance of microbial fuel cells. Bioresour Technol 2010;101:
[16] Walter XA, You J, Winfield J, Bajarunas U, Greenman J, Ieropoulos IA. From the 3520e5. https://doi.org/10.1016/j.biortech.2009.12.108.
lab to the field: self-stratifying microbial fuel cells stacks directly powering [45] Herbert-Guillou D, Tribollet B, Festy D. Influence of the hydrodynamics on the

9

ndez, M.J. Salar-García, D. Ruiz Ferna

n-Ferna
A. de Ramo
ndez et al. Energy 213 (2020) 118806

biofilm formation by mass transport analysis. Bioelectrochemistry 2001;53: 2008;1:487e96. https://doi.org/10.1111/j.1751-7915.2008.00049.x.

119e25. https://doi.org/10.1016/S0302-4598(00)00121-5. [47] Picioreanu C, van Loosdrecht MCM, Heijnen JJ. Two-dimensional model of
[46] Pham HT, Boon N, Aelterman P, Clauwaert P, De Schamphelaire L, Van biofilm detachment caused by internal stress from liquid flow. Biotechnol
Oostveldt P, et al. High shear enrichment improves the performance of the Bioeng 2001;72:205e18. https://doi.org/10.1002/1097-0290(20000120)72:
anodophilic microbial consortium in a microbial fuel cell. Microb Biotechnol 2<205::AID-BIT9>3.0.CO;2-L.

10