Content uploaded by Sujit Kumar
Author content
All content in this area was uploaded by Sujit Kumar on Apr 25, 2022
Content may be subject to copyright.
Backpropagation Algorithm-Based
Approach to Mitigate Soiling from PV
Module
Sujit Kumar and Vikramaditya Dave
Abstract Energy is viewed as a prime operator in the era of riches and a notewor-
thy figure of financial advancement. Constrained fossil assets and ecological issues
connected with them have underscored the requirement for new manageable energy
supply alternatives that utilize renewable energies. Among accessible innovations
for vitality generation from sun-based source, photovoltaic framework could give a
huge commitment to build up a more sustainable vitality framework. Utilization of
solar energy has not been opened up since the oil industry does not possess the sun.
The solar PV modules are for the most part utilized in dusty situations which is the
situation in subtropical nations like India. Dust gets gathered in the front surface of
the module and hinders the passage of light from the sun. It diminishes the power
era limit of the module. The power yield decreases as by half if the module is not
cleaned. Keeping in mind the end goal to routinely clean the dirt, a programmed
cleaning framework has been designed that both detects the dirt on solar panels and
cleans the surface of solar module consequently. The system consists of a panel
with wiper and water ejector. Artificial neural network (ANN) sends the signal to
the wiper motor with the help of measured data of solar irradiance, panel temper-
ature, PV voltage, current, and power in both sunny/cloudy as well as dusty days.
These data were well trained and tested along with the unknown data which were
not involved in the training. Backpropagation (BP) algorithm was used to train the
network, which showed the accuracy of 99% in data prediction and accordingly,
generated the control signal for windshield wiper motor and wiped the dust.
Keywords Solar energy ·Photovoltaic modules ·Neural networks ·
Backpropagation
S. Kumar (B
)
Noida Institute of Engineering and Technology, Greater Noida 201306, India
e-mail: sujitvj.kumar@gmail.com
V. D a v e
College of Technology and Engineering, Udaipur 313001, India
© Springer Nature Singapore Pte Ltd. 2020
A. Kalam et al. (eds.), Intelligent Computing Techniques for Smart Energy Systems,
Lecture Notes in Electrical Engineering 607,
https://doi.org/10.1007/978-981- 15-0214- 9_19
153
154 S. Kumar and V. Dave
1 Introduction
The severe augment in global warming and lubricate rates have convinced numerous
countries across the world to espouse novel energy policies that endorse renewable
energy application to ensure energy demand and to guard the surroundings. The
developed and developing countries are considering this fact and this has influenced
the energy sectors greatly. Solar energy is getting its due place and is used for various
applications which are replacing the cause of energy crisis. Moreover, solar energy
is natural, copious, gratis, fresh, and limitless.
PV cells technology makes solar power very effortless and effective. Till now, the
energy alteration proficiency of solar system has not reached a reasonable level so
the researchers find new scope to research in this field.
“Soiling” alluded to particulate pollution on the glass surfaces of PV module.
Ruining covers, soil gradual addition, sediment, salt, winged animal droppings, and
development of natural species which unfavorably influence the performance of
solar-based module.
Sand deposition is a multifaceted phenomenon and is prejudiced by various site-
specific ecological and climatic conditions. One of the major challenges before the
researchers is power loss caused by dust accumulation on the glass surfaces of solar
panels.
However, there has been less research work done over the removal of dust from
the solar panel. If adequate attention is put in the matter, the output efficiency will
be increased by a huge amount. The main motive of this paper is to model a basic
simulation model with adequate care to mitigate soiling from the PV panel using a
smart intelligent technique.
2 Factors Influencing Dust Settlement
A structure to appreciate the assorted aspects that drive the settling–absorption of dust
is adorned in Fig. 1. Globally soiling on solar panel surface is a huge problem. Sanaz
Ghazi [1] inspected the areas of dirt accumulation in various parts of the world and it
was proposed that the Middle East, North Africa, and some parts of Asia have the most
horrible dust buildup region in the world. Zaki Ahmad [2] considered the consequence
of dust contaminant type on panels. It was observed that diverse dust has diverse
characteristics, like red soil, residue, carbon; granite, and clay have more noteworthy
effect on PV. Juan Lopez-Garcia [3] studied the performance of PV modules by
examining different properties of durable soiling, which have been installed in the
open environment from last 30 years without washing in reasonable subtropical
weather. It was found that the flat glass modules display a higher difference in soiling
which weakens the competence of solar panel from 15 to 3%.
Solar power of two 1 MW PV plant was experimentally analyzed by Pavan Massi
[4]. It was found that soil type and the washing are the two techniques that are
Backpropagation Algorithm-Based Approach to Mitigate Soiling … 155
Fig. 1 Factors influencing dust settlement
strongly contingent on soiling. Loss of 6.9% was observed for the first plant built on
a quite sandy site due to pollution, whereas, loss of 1.1% was observed for the second
plant, constructed on an additional thick soil. Whitaker [5] examined the attainment
of solar panel and establish that the reduction of soiling phenomena strongly depend
on panel current and the voltage which depends on irradiance and array temperature.
The performance of selected PV system was simulated by Mayer [6] and con-
cluded that the PV module temperature and the in-plane large-scale irradiance influ-
ences the output power.
From the above-detailed survey, it is cleared that the settling of dirt on solar panels
diminish the efficiency based on the parameters like power, voltage, currents, array
temperature, and solar irradiance. Therefore, to overcome this discrepancy there must
be a cleaning method to precede the competence of the solar panels.
Halbhavi [7] designed an automatic cleaning system which senses the dust (using
light dependent resistor (LDR) sensor) on the solar panel and also cleans the module
automatically. This automated system is implemented using 8051 microcontroller
which controls the DC gear motor. Cleaning the PV modules was carried out by a
mechanism consisting of sliding brushes.
Cleaning techniques for PV surface has not focused on attention among the sci-
entists. Figure 2indicates diverse sorts of washing strategies which contain physical,
programmed, and uninvolved techniques.
Every technique has its benefits and drawbacks, for example, common practice
to clean windows of buildings is a labor-intensive (manual cleaning). A few brushes
specifically fitted to a water supply to play out the washing and scouring at the same
time. Unlikely, a stepping stool and a cleaner with long handle are expected to wash
the board which is again a chaotic schedule and work increases. In the present time,
artificial intelligence (AI) is one of the most acclaimed methods which can resolve
the above issue.
156 S. Kumar and V. Dave
Fig. 2 Cleaning techniques adopted for expelling dust from solar panel
3 Training and Modeling of ANN
Transformative computations such as artificial neural networks (ANN), fuzzy sys-
tems, and their combination, along with other general machine learning methods
and profound learning methods, e.g., clustering, classification, knowledge-based
systems, case-based reasoning, decision-making methods, etc. are used as exten-
sive range of AI techniques. Among these, ANN is being used from last 60 years, to
grip practical problems, application software has been developed in early 30 years.
Artificial neural network (ANN) controller is a standout among the most critical
strategies for the estimation of non-straight frameworks [8,9].
ANN utilizes distinctive training algorithms like backpropagation (BP), Leven-
berg Marquardt (LM), radial basis function (RBF), etc. Among these BP algorithm is
superior to other algorithms due to its less union time and precision. Fundamentally,
the backpropagation neural network (BPNN) is a multilayer perceptron network
(MLP) with error backpropagation algorithms [10]. BP algorithm comprise of a
three-layer feed-forward perceptron neural network architecture which has an input,
a hidden, and an output layer and each layer has several independent neurons as (1). In
particular, BP-NN requires the activation function to be continuously differentiable
The three stages of network, learning process are shown below [10]:
•Forward propagation stage. It is corresponding to MLP and develops a set of output
signal based on the input.
•Backpropagation error stage. Initially, the error between the output signal and the
ideal value in the output layer is calculated. Then the error propagates backward
layer by layer.
•Update weight stage. The weight matrix of each layer is attuned based on the error
which is in the accordance of gradient descent principle.
Backpropagation Algorithm-Based Approach to Mitigate Soiling … 157
Figure 3demonstrates the ANN model of the total framework system which
comprises PV module associated with a load of 200 W. In the present experiment,
data from the solar panel has been measured in the form of panel current, voltage,
temperature, solar irradiance, and power in bright and hazy days with and without
dirt. Panel consists of windshield wiper with a small jet which injects water at the
time of wiping the dust from the panel. The supply to the windshield wiper motor
gets from the solar panel itself after being getting a control signal from the ANN. An
example has been illustrated in Table 1. The system gets the exterior data, calibrates
Fig. 3 Schematic diagram of ANN controlled PV system
Table 1 Data of different types of days measured from solar panels
Type of
day
Time
(a.m/p.m)
Vo l t a g e
(V)
Currents
(A)
Power (W) Solar
irradiance
(W/m2)
Module
tempera-
ture
(°)
Bright 12:00 p.m 18.71 7.6 142.2 764 48.68
Bright 1:00 p.m 22.83 8.1 185 1000 57.50
Bright 11.30 a.m 13.85 6.6 91.47 465 43.10
hazy 11:00 a.m 9.39 3.1 29.125 151 24.40
hazy 12:00 p.m 13.10 2.9 38 199 27.20
hazy 11:30 a.m 12.6 4.5 56.72 276 31.20
Dusty
bright
11:00 a.m 12.97 2.59 33.65 564 46.10
Dusty
bright
12:00 p.m 13.5 3.15 42.63 793 46.40
Dusty
hazy
12:30 p.m 10.97 2.80 30.72 276 39.40
158 S. Kumar and V. Dave
it by weights and relays it to following layer with the help of neurons. Tan sigmoid
activation function is being used for input and hidden layer neurons while pure linear
activation function is being used for output layer neurons.
Input to the ANN are PV array parameters; PV voltages (Vpv), currents (Ipv)
and power (P) and environmental data: irradiance (G) and module temperature (T)
(comprises five data) which is then trained to evaluate the performance of the ANN
model. The output layer has a single-output node which is either one or zero. One
denotes the dust is collected over the module which then produces the signal to wiper
motor to wash the panel and zero denotes no settlement of dirt over the module.
4 Results and Discussion
Miscellaneous models are assessed and the best one is picked by experimentation.
The best option comprises 5 neurons in the input layer, 30 in the hidden layer, and
1 neuron in the output layer as illustrated in Fig. 4. The following sub-database was
extracted from the data set described above.
– Inputs to the ANN were (T, G, P, Vpv, and Ipv) arranged to meet the circumstance
as standard; each day solar emission is lower or higher than 1000 W/m2/day (dark
or clear days with and without dirt). There were 5365 samples as sub-database.
In training stage, to solve over the fitting problem 10% cross-validation method
is used, whereas, for training the network 80% of samples used, 10% utilized to
validate the network, and 15% utilized for test process.
To equitably assess the performance of the systems, four diverse factual markers
were utilized. These markers are mean absolute error (MAE), mean squared error
Fig. 4 Projected architecture for ANN
Backpropagation Algorithm-Based Approach to Mitigate Soiling … 159
(MSE), and mean absolute percentage error (MAPE):
MAE =
1
n
n
i=1
Ypredicted −Ytrue
(1)
MSE =
1
n
n
i=1
Ypredicted −Ytrue2(2)
MAPE =
1
n
n
i=1
Ypredicted −Ytrue
Ytrue
2
(3)
where Ypredicted and Ytrue are estimated and measured dust values over panels by
the models, respectively. Among the above factual measures, MAPE is the most
critical measurable quantity in that it mentions utilization of every single objective
reality and has the smallest fluctuation from sample to sample. Variety of users
easily understands MAPE, so it is frequently utilized for reporting [11]. However,
MSE will also be used for performance analysis as based on this the optimum number
of neurons in hidden layer will be decided. Here, MSE parameter is used to stop the
training process and results are shown in Fig. 5.
The accompanying parameters are placed through training our ANN model:
Training pattern =5365 samples, learning rate =0.001, set error objective =
0.01, number of epochs =100, momentum =0.95.
Fig. 5 Training results of ANN with evolution of the performance
160 S. Kumar and V. Dave
Fig. 6 Test phase
performance of ANN model
The ANN generates the signal only when the panel is accumulated by dust as the
data provided in the input was having dust data in both dusty and cloudy days. ANN
model was tested with the other data which were not included in the training set to
ensure that the prediction of dust by the ANN is fulfilling or not. Mean square error
(execution goal =9.05 ×10−2) is achieved which is well below the set objective in
100 epochs as shown in Fig. 5. The validation showed the accuracy of trained ANN
model uses other data which were not used in the training and the mean square error
observed for the validation was 8.4968 ×10−3which was less than the set error
objective.
Figure 6shows results of tested network using whole dataset. 1.00052e−3is
obtained as network test error and 9.98791e−1 is obtained as regression coefficient
(R). Regression coefficient tells about how well our data has been classified and fit
to the actual data. MAPE of the ANN model is found to be below 0.5% and in this
way taking into report the accuracy in the estimation of solar irradiance, module
temperature, current, voltage, and misfortunes in association wires, which is roughly
0.5%, consequently, on the whole error is around 1%.
5 Conclusion
Automatic washing with smart intelligent technology proved to be more economic
and significantly less cumbersome when compared to manual operation, particularly
in systems having large number of solar panels. Trouble-free ANN model has been
simulated that generates the signal to the motor operating wiper over the occurrence
of dust over solar panel of 250 Wp polycrystalline PV modules.
From the results presented, it can be concluded that ANN has capability to effec-
tively predict the dust data with an accuracy of 99% and successfully operated the
Backpropagation Algorithm-Based Approach to Mitigate Soiling … 161
wiper which wipes the panel. The approval too included investigation of the ade-
quacy of the trained data when copying with obscure information, i.e., data which
are not included in the training. This proves the system has high robustness when
compared to other techniques.
References
1. Ghazi S, Sayigh A, Ip K (2014) Dust effect on flat surfaces—a review paper. Renew Sustain
Energy Rev 742–751
2. Darwish ZA, Kazem HA, Sopian K, Al-Goul M, Alawadhi H (2015) Effect of dust pollutant
type on photovoltaic performance. Renew Sustain Energy Rev 735–744
3. Lopez-Garcia J, Pozza A, Sample T (2016) Long-term soiling of silicon PV modules in a
moderate subtropical climate. Sol Energy 174–183
4. Massi Pavan A, Mellit A, De Pieri D (2011) The effect of soiling on energy production for
large-scale photovoltaic plants. Sol Energy 1128–1136
5. Whitaker CM, Timothy U, Townsend TU, Newmiller JD, King DL, Boyson WE, Kratochvil
JA, Collier DE, Osborn DE (1997) Application and validation of a new PV performance char-
acterization method. In: Twenty-sixth IEEE photovoltaics specialist conference and exhibition,
Anaheim, California, USA, pp 1253–1256
6. Mayer D, Wald L, Poissant Y, Pelland S (2008) Performance prediction of grid-connected
photovoltaic systems using remote sensing. Report IEAPVPST, pp 18–26
7. Halbhavi SB, Kulkarni SG, Kulkarni DB (2015) Microcontroller based automatic cleaning of
solar panel. Int J Latest Trends Eng Technol (IJLTET) 99–103
8. Dzung PQ, Phuong LM, Vinh PQ, Van Nho N, Hien DM (2006) The development of artificial
neural network space vector PWM and diagnostic controller for voltage source inverter. In:
IEEE power India conference, New Delhi, India, pp 10–12
9. Dzung PQ, Phuong LM, Vinh PQ (2007) The development of artificial neural network space
vector PWM for four-switch three-phase inverter. In: International conference on power elec-
tronics and drive systems—IEEE PEDS 2007, Thailand, 27–30 Nov 2007
10. Cheng T-R, Li Y (2016) Research status of artificial neural network and its application assump-
tion in aviation. In: 12th international conference on computational intelligence and security,
pp 407–410
11. Celik AN (2007) Artificial neural network modeling and experimental verification of the oper-
ating current of mono-crystalline photovoltaic modules. Sol Energy 2507–2517
