Hindawi

International Journal of Photoenergy

Volume 2022, Article ID 4921153, 6 pages
https://doi.org/10.1155/2022/4921153

Research Article
Improving the Efficiency of Photovoltaic Panels Using Machine
Learning Approach

Rashmita Khilar,1 G. Merlin Suba,2 T. Sathesh Kumar,3 J. Samson Isaac,4

Santaji Krishna Shinde,5 S. Ramya,6 V. Prabhu,7 and Kuma Gowwomsa Erko 8

1
Department of Information Technology, Saveetha School of Engineering, Saveetha Institute of Medical and Technical Sciences,
Saveetha University, Chennai, Tamil Nadu 600124, India
2
Department of Electrical and Electronics Engineering, Panimalar Engineering College, Poonamallee, Chennai,
Tamil Nadu 600123, India
3
Department of Electrical and Electronics Engineering, Dr. Mahalingam College of Engineering and Technology, Pollachi,
Tamil Nadu 642003, India
4
Department of Biomedical Engineering, Karunya Institute of Technology and Sciences, Coimbatore 641114, India
5
Department of Computer Engineering, Vidya Pratishthan’s Kamalnayan Bajaj Institute of Engineering and Technology, Baramati,
Maharashtra 413133, India
6
Department of Information Technology, M.Kumarasamy College of Engineering, Karur, Tamil Nadu 639113, India
7
Department of Mechanical Engineering, Sri Sairam Engineering College, Chennai, Tamil Nadu 600044, India
8
Department of Mechanical Engineering, Ambo University, Ethiopia

Correspondence should be addressed to Kuma Gowwomsa Erko; kuma.gowwomsa@ambou.edu.et

Received 29 January 2022; Revised 10 April 2022; Accepted 15 April 2022; Published 28 May 2022

Academic Editor: V. Mohanavel

Copyright © 2022 Rashmita Khilar et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

Photovoltaic (PV) solar panels account for a major portion of the smart grid capacity. On the other hand, the accumulation of
solar panels dust is a significant challenge for PV-based systems. The accumulation of solar panels dust results in a significant
reduction in the amount of energy produced. Because of the country’s low wind velocity and rainfall, frequent cleaning of solar
panels is necessary either by manual or automated means. Cleaning activities should only be initiated when absolutely essential
to reduce maintenance costs and increase the power output of solar panels that have been projected to be affected by dust
accumulation. In this paper, we develop a deep belief network model to detect the dust particles in the solar panels installed as
a large unit. The study takes into account various input metrics that includes solar irradiance, temperature level, and dust level
on the panels. These metrics are used for the estimation of the level of dust present in the atmosphere and how often the
panels can be cleaned at regular intervals. The simulation is conducted to test the efficacy of the model in cleaning the panels.
The results are estimated in terms of accuracy, precision, recall, and F-measure. The results of the simulation show that the
proposed model achieves higher accuracy rate of more than 99% than other methods.

1. Introduction energy sources into power systems has become a global imper-
ative. Solar, wind, biomass, geothermal, and fuel cells are just a
The use of thermal power plants contributes to the accelera- handful of the renewable energy sources that can be used to
tion of global warming. In contrast, the fossil fuels that are power electric grids and other energy systems.
used in these plants are getting increasingly expensive and dif- The demand for natural gas and oil tends to increase
ficult to come by. As a result, the integration of renewable dramatically in recent years, causing countries to look for
2 International Journal of Photoenergy

alternative energy sources. Solar power and renewable energy Solar

technology are being actively considered as viable alternative photovoltaic
energy sources and other countries across the world. With its surface
favourable temperature and geographical location, it is home
to the world’s highest solar power generation. As a result, solar Dust settled
energy has become the most widely used renewable energy
source.
Photovoltaic panels are made with semiconductor mate-
Dust flying
rials, which are employed in the manufacturing process. The
inherent properties of photovoltaic materials have an impact
on the efficiency of photovoltaic systems. Solar panel output Solar
is therefore dependent on the materials used in their manu- irradiance
facture as well as the coating applied to the glass on which
they are constructed.
As a result, several impediments to the widespread use of Water level
present
solar energy in underdeveloped countries. The impact of dust
is one of the most significant considerations. Because of the
country’s geographic location, the pace at which dust accumu-
Temperature
lates is extremely high. A build up of dust on photovoltaic
panels causes them to degrade quickly, which results in a
significant reduction in their power production. Both the envi-
ronmental conditions and the qualities of the dust play a role Rainfall
in determining the amount of dust that accumulates in a given
area. The weight, shape, and size of the dust particles are all Figure 1: Main sources of data collection on PV panels.
characteristics of the dust, whereas environmental elements
include the weather and the particle location [1–3]. The main contribution of the paper involves the following:
Additionally, the wind might have an impact on dust set-
tling. A large amount of dust collects on the surface of solar (i) The study takes into account various input metrics
panels in regions, causing the power output of solar photo- that includes solar irradiance, temperature level,
voltaic panels to decrease at the fastest rate. Climate factors and dust level on the panels
such as the temperature and humidity of the surrounding
air have an impact on dust settlement. Solar panels lose their (ii) These metrics are used for the estimation of the level
efficiency as a result of increased dust accumulation on them of dust present in the atmosphere and how often the
as the temperature and humidity rise [4]. panels can be cleaned at regular intervals
As a result, solar panels should be cleaned on a regular
basis to ensure maximum efficiency. However, if no cleaning
procedures are required, this strategy is not cost-effective 2. Background
and may result in a waste of resources [5]. It is critical to
determine the quantity of dust on solar panels and to begin According to the research literature, solar photovoltaic pro-
cleaning procedures as soon as possible in order to maximise duction can be investigated either via the use of tests or
output while reducing maintenance costs. The impact of through the development of prediction models.
dust collected on panels varies depending on the weather In [6], the effects of dust, humidity, and air velocity were
conditions and the size of the dust particles. The output investigated individually and in combination. Fine particles
power of PV panels also fluctuates depending on the loca- have a greater impact on the efficiency of solar panels. A
tion, the properties of the dust, and the temperature of the greater tilt angle also helped to prevent dust accumulation,
environment. As a result of this limitation, the outcomes of albeit wetness accelerated the coagulation of dust particles.
research undertaken in one country cannot be generalised Tests were carried out in [7] in order to demonstrate the
to other countries. effect of pollutant type and weight on the output power of
It is feasible to employ the process described in this solar-powered PV panels.
paper in other countries, but not the data or findings. As a In [8], it was found that three unique fake contaminants
result, it is vital to continuously monitor the quantity of dust were injected into the solar photovoltaic panel during the
in photovoltaic farms and to take appropriate steps to reduce experiment. The testing included the use of pollutants as con-
maintenance costs while simultaneously boosting output. In taminants. The results revealed that solar panels produced
fact, the generality of the methodology given herein distin- substantially more electricity than clean panels. A comparable
guishes it from other approaches as it can be used to detect study into three artificial pollutants, red soil, sand, and ash,
dust accumulation regardless of where it occurs. was conducted and the results were published in [9].
In this paper, we develop a deep belief network model to A 23-day test of self-cleaning coatings on solar panels was
detect the dust particles in the solar panels installed as a conducted [10] to see whether they were effective. When com-
large unit as in Figure 1. pared to panels that did not have this coating compound, the
International Journal of Photoenergy 3

performance of the panels with this coating compound was 3. Proposed Method
practically equal.
A number of experiments were carried to investigate the Forecasting solar panel dust accumulation accurately is cru-
dust impact on a solar panel performance. The dust concen- cial for investors and grid operators alike, as it affects their
trations in solar panel modules ranged from 0.0063 g/m2 to bottom lines. As shown in Figure 2, the research was aimed
0.36 g/m2, depending on the model. According to the at estimating the dust level with various measurements on
research, a linear link was discovered between dust mass solar panels. If there is a significant amount of dust, cleaning
and the fall in solar output power. According to [11], satel- procedures may be initiated. Machine learning was used in
lite images and an SVM model were used to estimate cloud the development of this detector. The first step in developing
movement and irradiance. The autoregression model pre- a machine learning model for photovoltaic panels is to col-
sented in [12] was used to forecast the output over a period lect all of the data that will be needed. After that, a number
of up to 36 hours, and the results were positive. of regression models are constructed and tested on the data
As reported in [13], the ANN technique provides predic- that has been collected.
tions of global solar radiation based on data from weather
stations throughout the world. The hourly diffuse solar irra- 3.1. Data Preparation. Analysis of data before it is fed into
diance was estimated using a sigmoid function regression the model is used to predict dust levels, which is accom-
model by the authors in [14] for all weather conditions. plished through the use of regression models. The statistical
The model discovered a 25–35 percent relative root RMSE. approach helps to predict the relationship between the
The output power of a wind turbine in Turkey may be calcu- response and single output variable, and one or more input
lated using an ANN model over a variety of time periods and variables, referred to as predictors; before entering the
throughout various seasons [13]. A solar panel with a capac- model, the predictors and response variables should be
ity of 750 watts was installed (PV). The RMSE for each sea- grouped in a matrix to make it easier to understand. In this
son and time period was tallied separately. study, three factors are used to predict the value of a single
In [15], the output power of California PV panels was pre- output or response variable. When operating in these condi-
dicted using four different approaches, one of which being the tions, PV power output is influenced by three variables: the
ANN method. It was decided to use a one-megawatt solar amount of solar irradiation received, the temperature of
panel field to forecast one hour and two hours in the future. the surrounding air, and a reaction variable termed dust. A
In terms of accuracy, the ANN surpassed all of the other fore- comprehensive collection of data was acquired through a
casting methods except for one. According to the research, series of tests, which will be discussed in further detail in
there was an inaccuracy of up to 20% in the other models. In the following paragraphs and which are conducted over a
the provided model, the root mean square error (RMSE) was number of days, resulting in an extensive collection of data.
found to be 15%. Neither the study nor the literature made
any considerations for how environmental circumstances 3.2. Model Training. After that, the model is fed with data
would affect the output power of the system. from the tests, which is then trained on that data to teach
Using solar modules as an example, researchers in Spain it how to predict dust levels based on inputs such as the
conducted a study to determine how rainfall impacts the power output of PV panels, the temperature of the environ-
cleaning of the modules [16]. According to a study, solar pho- ment, and the amount of solar irradiation. There are two
tovoltaic power plants lose 20 percent of their energy during types of data collected (predictor variables and responses).
dry seasons, compared to only 4.4 percent during wet months, Data sets are used in the model creation and validation pro-
due to dust accumulation on the solar panel surfaces [12]. cess for both the training and testing stages. Data sets are
As part of another study experiment in Morocco, the used in both the training and testing stages.
output of a solar photovoltaic panel and rainfall was mea- The solar irradiance, temperature, and output power are
sured for four months and then used to determine the accu- all included in the regression models as three predictors and
mulation of dust. The information on the amount of rain one response variable. The regression model predicts the
that fell came from a weather data collection centre. dust present, which is then compared to the actual response.
Studies in [17] looked into how dust collecting affects PV While training, the 5-fold cross-validation method is used. It
panel performance in the Jazan region, despite the fact that is often used in regression models to analyse and quantify
rain occurs seldom and with a low degree of intensity. The effi- misclassification errors, which is a common occurrence.
ciency of photovoltaic panels was found to be reduced by 10% During the training phase of this method, five sets of data
as a result of the frequent dust collection. Lower tilt angles, on are picked at random from the training data set to be used
the other hand, gathered far more dust than higher tilt angles. in the training phase.
The authors of this article were inspired to write it because The four groups will be used to train and evaluate the
they wanted to share their knowledge with others. model during the development process. This technique will
In [18], indications are conducted in order to determine be repeated numerous times in order to ensure that each group
the dust accumulated on photovoltaic surface. Much research has been examined at least once before moving on. Following
has been conducted to determine the dust impact on photo- the training process, the difference between the actual and pro-
voltaic panels. There have been a few studies conducted to jected response is found determined using the data collected.
determine the amount of dust that has to be eliminated before When using different models, it is feasible to estimate solar
the proper cleaning procedure can begin. panel dust collection with a different standard deviation
4 International Journal of Photoenergy

Temperature Irradiance Dust level l1 l2 lj lL

Estimate error

h1 h2 hj hJ 1
Panels
Feature extraction

h1 h2 hj hJ 1
Data preparation Classification

h1 h2 hj hJ 1

Predicted dust level

v1 v2 vi vI 1
Figure 2: Proposed algorithm.
Figure 3: DBN/DNN architecture.
(RMSE). From among the available regression prediction
models, the one that provides the most accurate dust estima- number of hidden layers has been achieved, a complete
tion with the lowest RMSE is chosen. The correctness of a backpropagation fine tuning procedure is carried out. Using
model can be tested by using a variety of case studies after it this pretraining strategy has been proven to be effective on
has been narrowed down to the most accurate one through a the battlefield.
process of elimination.
When RBMs are trained, they are stacked together to pro- 3.3. Dust Estimation Unit. After the model selection with the
duce a DBN. The following is an explanation of the stacking lowest root mean square error (RMSE), this regression model
technique. Using the activation probabilities of its hidden is extracted. The temperature, output power, and solar irradi-
units, a Bernoulli-Bernoulli RBM is trained after learning a ance are fed into the system, and it can anticipate the amount
Gaussian-Bernoulli RBM or a Bernoulli RBM, which is then of dust that will build up on the solar panel in the future. Those
used to train another RBM. After that, the activation probabil- in charge of operating solar photovoltaic systems rely on
ities of the second-layer Bernoulli-Bernoulli RBM are used as precise dust estimates to decide the optimal time to clean the
the visible data input by the third-layer Bernoulli-Bernoulli modules. When the dust level on solar PV panels surpasses a
RBM. This is a three-layer RBM with three layers. The layer- specified threshold, the amount of energy produced by the sys-
by-layer greedy learning improves the training data under tem can be considerably reduced. As a result, as soon as the
the stacking, which provides some theoretical evidence for this level of dust collection exceeds this threshold, cleaning activities
effective layer-by-layer greedy learning technique. Thus, the can commence. Solar photovoltaic modules should be cleaned
greedy strategy described above achieves a close approxima- at the most opportune time of year to save on the extra costs
tion to the maximum probability distribution. It is vital to associated with cleaning solar photovoltaic modules on a regu-
remember that this learning process is unsupervised and does lar basis. In the final stage, a new data set is used to test the dust
not necessitate the use of a class label. estimation unit, which helps to minimise bias and evaluate
Different learning approaches, often discriminative, that accurately the performance by evaluating its accuracy.
fine-tune the weights together to improve the network
performance in classification tasks are used in addition to 4. Results and Discussions
pre-training. After that, variables reflecting the outputs from
the training data are introduced to further fine-tune the per- It has already been stated that a large amount of data is
formance and accuracy of the discriminative model. required for both training and testing the model. Using a
The back-propagation fine-tunes the weights of a feed- solar photovoltaic panel in the open air under realistic envi-
forward neural network. What appears in the top layer is ronmental circumstances, experiments were conducted to
determined in this case. HMM-based speech recognition determine its performance. Researchers have looked into
systems employ the top layer, shown in Figure 3 by the letters the impact of dust on the performance of photovoltaic
l1, l2,… lj, lL, to represent syllables, phone calls, subphones, or panels in these types of environments. The scientists began
other speech units, while the lower layers represent additional by setting up a 400-watt solar photovoltaic panel as a test
speech units. bed for their experiments.
Phone and speech detection were among the tasks on The dust that was used in the experiment was collected
which the generative pretraining described above outper- from the same location where the experiment was con-
formed random initialization across a broad range of tasks. ducted. The output power of the solar PV panel was tested
In addition, the order in which minibatches are formed must under a variety of conditions, including different times of
be carefully studied prior to being implemented. As a result, the day, varying degrees of solar irradiation, different tem-
it was determined that starting with a simple neural network peratures, and different dust concentrations. For this exper-
with one hidden layer and working your way up was the iment, the RS PRO Solar Power Meter ISM400 was utilised
most successful way to learn the DNN. Once the necessary to measure solar irradiance.
International Journal of Photoenergy 5

Table 1: Accuracy. proposed model achieves higher precision rate with its effi-
cient computation.
Panels ANN BPNN Proposed Table 3 shows the results of recall between the proposed
Series parallel 0.94209 0.94432 0.95687 model and existing machine learning models. The results of
Parallel 0.94198 0.94357 0.95683 simulation on various PV panels show that the proposed model
Bridge link 0.94117 0.94355 0.95677 achieves higher recall rate with its efficient computation.
Honeycomb 0.94088 0.94285 0.95664 Table 4 shows the results of F-measure between the pro-
posed model and existing machine learning models. The
Total cross tied 0.93368 0.94280 0.95641
results of simulation on various PV panels show that the
proposed model achieves higher F-measure rate with its effi-
Table 2: Precision. cient computation.
Table 5 shows the results of MAPE between the proposed
Panels ANN BPNN Proposed model and existing machine learning models. The results of
Series parallel 0.87254 0.87307 0.89406 simulation on various PV panels show that the proposed
Parallel 0.85895 0.86283 0.89200 model achieves reduced MAPE with its efficient computation.
Bridge link 0.85859 0.86281 0.88695
Honeycomb 0.83920 0.84169 0.88624 5. Conclusions
Total cross tied 0.83801 0.84148 0.87110
In this paper, deep belief network model is used to detect the
dust particles in the solar panels installed as a large unit. The
Table 3: Recall. study takes into account various input metrics that includes
solar irradiance, temperature level, and dust level on the
Panels ANN BPNN Proposed panels. These metrics are used for the estimation of the level
Series parallel 0.78337 0.83903 0.94382 of dust present in the atmosphere and how often the panels
Parallel 0.83004 0.84054 0.94434 can be cleaned at regular intervals. The simulation is con-
Bridge link 0.83978 0.84067 0.95362 ducted to test the efficacy of the model in cleaning the panels.
Honeycomb 0.85972 0.85202 0.95580 The results are estimated in terms of accuracy, precision,
recall, and F-measure. The results of the simulation show that
Total cross tied 0.86925 0.90605 0.95641
the proposed model achieves higher accuracy rate than other
methods. In the future, deep learning models can be used to
Table 4: F-measure. improve the rate of accuracy.
Panels ANN BPNN Proposed
Data Availability
Series parallel 0.95449 0.90118 0.94118
Parallel 0.95174 0.89384 0.93457 The data used to support the findings of this study are
Bridge link 0.94593 0.89250 0.93276 included within the article. Further data or information is
Honeycomb 0.93495 0.88444 0.93180 available from the corresponding author upon request.
Total cross tied 0.86860 0.88444 0.89673
Conflicts of Interest
Table 5: MAPE. The authors declare that there is no conflict of interest
regarding the publication of this paper.
Panels ANN BPNN Proposed
Series parallel 0.08574 0.01366 0.00311
Acknowledgments
Parallel 0.18026 0.02100 0.00201
Bridge link 0.20701 0.02234 0.00192 The authors thank the Ambo University, Ethiopia, for pro-
Honeycomb 0.24857 0.04624 0.00304 viding help during the research and preparation of the man-
Total cross tied 0.23473 0.04091 0.00304 uscript. The author thank the Saveetha University, Vidya
Pratishthan’s Kamalnayan Bajaj Institute of Engineering
and Technology, Dr. Mahalingam College of Engineering
and Technology, M. Kumarasamy College of Engineering,
Table 1 shows the results of accuracy between the pro- and Sri Sairam Engineering College for providing assistance
posed model and existing machine learning models. The in completing the work.
results of simulation on various PV panels show that the
proposed model achieves higher accuracy rate with its effi-
References
cient computation.
Table 2 shows the results of precision between the pro- [1] W. A. Hanafy, A. Pina, and S. A. Salem, “Machine Learning
posed model and existing machine learning models. The Approach for Photovoltaic Panels Cleanliness Detection,” in
results of simulation on various PV panels show that the In 2019 15th International Computer Engineering Conference
6 International Journal of Photoenergy

(ICENCO), pp. 72–77, New Jersey, United States, December [16] R. Usamentiaga, A. Fernández, and J. L. Carús, “Evaluation of
2019. dust deposition on parabolic trough collectors in the visible
[2] K. Priyadharsini, A. Srikanth, V. Sounddar, and and infrared spectrum,” Sensors, vol. 20, no. 21, p. 6249, 2020.
M. Senthamilselvan, “Elegant method to improve the effi- [17] K. Benhmed, F. Touati, M. Al-Hitmi et al., “PV Power Predic-
ciency of remotely located solar panels using IoT,” Materials tion in Qatar Based on Machine Learning Approach,” in 2018
Today: Proceedings, vol. 45, pp. 8094–8104, 2021. 6th International Renewable and Sustainable Energy Confer-
[3] P. Sivagami and N. M. Jothiswaroopan, “IOT based statistical ence (IRSEC), pp. 1–4, New Jersey, United States, December
performance improvement technique on the power output of 2018.
photovoltaic system,” Journal of Ambient Intelligence and [18] B. Yılmaz and R. Yıldırım, “Critical review of machine learning
Humanized Computing, vol. 12, no. 5, pp. 5029–5043, 2021. applications in perovskite solar research,” Nano Energy,
[4] M. Heinrich, S. Meunier, A. Same et al., “Detection of cleaning vol. 80, article 105546, 2021.
interventions on photovoltaic modules with machine learn-
ing,” Applied Energy, vol. 263, article 114642, 2020.
[5] A. Khandakar, M. EH Chowdhury, M. Khoda Kazi et al.,
“Machine learning based photovoltaics (PV) power prediction
using different environmental parameters of Qatar,” Energies,
vol. 12, no. 14, article 2782, 2019.
[6] A. J. Trappey, P. P. Chen, C. V. Trappey, and L. Ma, “A
machine learning approach for solar power technology review
and patent evolution analysis,” Applied Sciences, vol. 9, no. 7,
article 1478, 2019.
[7] C. G. Haba, “Monitoring Solar Panels Using Machine Learn-
ing Techniques,” in 2019 8th International Conference on
Modern Power Systems (MPS), pp. 1–6, New Jersey, United
States, May 2019.
[8] Z. Ebrahimi-Khusfi, A. R. Nafarzadegan, and F. Dargahian,
“Predicting the number of dusty days around the desert wet-
lands in southeastern Iran using feature selection and machine
learning techniques,” Ecological Indicators, vol. 125, article
107499, 2021.
[9] M. K. Swain, M. Mishra, R. C. Bansal, and S. Hasan, “A self-
powered solar panel automated cleaning system: design and
testing analysis,” Electric Power Components and Systems,
vol. 49, no. 3, pp. 308–320, 2021.
[10] G. M. Dantas, O. L. C. Mendes, S. M. Maia, and A. R. de Alex-
andria, “Dust detection in solar panel using image processing
techniques: a review,” Research, Society and Development,
vol. 9, no. 8, pp. e321985107–e321985107, 2020.
[11] M. Altıntaş and S. Arslan, “The study of dust removal using
electrostatic cleaning system for solar panels,” Sustainability,
vol. 13, no. 16, p. 9454, 2021.
[12] Y. Natarajan, S. Kannan, C. Selvaraj, and S. N. Mohanty,
“Forecasting energy generation in large photovoltaic plants
using radial belief neural network,” Sustainable Computing:
Informatics and Systems, vol. 31, article 100578, 2021.
[13] Z. Ebrahimi-Khusfi, R. Taghizadeh-Mehrjardi, and
M. Mirakbari, “Evaluation of machine learning models for
predicting the temporal variations of dust storm index in arid
regions of Iran,” Atmospheric Pollution Research, vol. 12, no. 1,
pp. 134–147, 2021.
[14] B. Gobinathan, M. A. Mukunthan, S. Surendran et al., “A novel
method to solve real time security issues in software industry
using advanced cryptographic techniques,” Scientific Program-
ming, vol. 2021, 9 pages, 2021.
[15] I. M. Kirpichnikova and V. Shestakova, “Problems of Using
Solar Photovoltaic Panels and Ways to Increase their Effi-
ciency,” in 2020 International Conference on Industrial Engi-
neering, Applications and Manufacturing (ICIEAM), pp. 1–7,
New Jersey, United States, May 2020.