Kegel Eers

FACULTY OF ENGINEERING TECHNOLOGY
TECHNOLOGY CAMPUS DE NAYER
The development of a clean-

ing robot for PV panels
Matthias KEGELEERS
Promotor: Prof. Dr. Ir. T. Goedeme Master Thesis submitted to obtain the degree of
Master of Science in Engineering Technology:
Co-promotors: Prof. Dr. P. He Electromechanics, Manufacturing Engineering
Ir. S. Landkammer
Academic Year 2014 - 2015

c Copyright KU Leuven
Without written permission of the supervisor(s) and the author(s) it is forbidden to reproduce or
adapt in any form or by any means any part of this publication. Requests for obtaining the right
to reproduce or utilise parts of this publication should be addressed to KU Leuven, Technology
Campus De Nayer, Jan De Nayerlaan 5, B-2860 Sint-Katelijne-Waver, +32 15 31 69 44 or via email
fet.denayer@kuleuven.be.
A written permission of the supervisor(s) is also required to use the methods, products, schemat-
ics and programs described in this work for industrial or commercial use, and for submitting this
publication in scientific contests.
Acknowledgements
Foremost, I would like to express my sincere gratitude to my promotor Prof. Toon Goedeme from
the K.U. Leuven for the support on working out subject for this thesis and the guidance during the
project.
Also, I would like to thank my co-promotors Prof. Peter He and Ir. Stefan Landkammer from the
Technische Hochschule Nurnberg for their support during my stay at their faculty and for helping
me with the practical realisations of the thesis.
My sincere thanks also go to Peter Arras and Anastassia Knor, international coordinators for re-
spectively KU Leuven and TH Nurnberg for making it possible to have the experience of studying a
semester abroad.
I also want to thank my fellow students Tim Van Assche and Toon Witters for helping me with staying
into contact with the faculty in Belgium and passing along information about the practicalities of the
paper. Also I thank Koen Dumortier for helping me with printing and delivering my thesis.
iii
Abstract (English)
Growing interest in renewable energy has led the solar photovoltaic industry to expand notably in
the last decade. However a big issue that is often overlooked too easily, is keeping the panels
clean. Panels are often difficult and dangerous to reach and not worth the risk cleaning them.
Furthermore, autonomous cleaning robots are often only economical on a larger scale due to both
installation costs and the fact that custom-made parts are needed to fit the plant.
This papers focus is on finding a more cost-worthy solution for the drive principle in order to de-
crease the overall price of the robot, thus making it profitable on smaller scaled PV-plants. The goal
is for it to return its investment in two years (being the warranty period in most European countries)
while keeping an average family sized plant clean.
A principle, were a robot uses two wires attached to the edges of the roof, was investigated. This
driving principle was virtually modelled to simulate its performance. Afterwards a proof of concept
was built to validate the model.
The research found that this drive principle is a promising alternative when applied to small plants.
The system is both agile, flexible and very cost effective. According to a rough estimation, for an
average family, the robot would earn its price back in two years at losses of 15%. Bigger sized
plants return their investment at even lower losses. The reliability has to be improved though.
iv
Abstract (Dutch)
Door de groeiende interesse in hernieuwbare energie is de vraag naar fotovoltasche zonnepanelen

sterk gestegen door de marktgroei het laatste decennium. Tevens is ook het onderzoek gestegen
naar het verbeteren van de efficientie van de panelen. Een groter probleem wordt echter vaak over
het hoofd gezien. Het reinigen van de panelen zorgt voor een verbeterde efficientie. Dit is echter
niet zo eenvoudig omdat het vaak moeilijk en gevaarlijk is de panelen te bereiken.
Daarnaast zijn de autonome robots die op heden gebruikt worden voor het reinigen van PV cellen
vaak niet rendabel op een kleine schaal. De robots moeten vaak op maat gemaakt worden en ook
de installatie ervan kost extra geld.
Deze paper focust op het vinden van een alternatieve aandrijving die de prijs kan drukken waardoor
het ook financieel verantwoordbaar is op kleine schaal toe te passen. Het doel is om een robot te
ontwikkelen die zichzelf in twee jaar terugverdient (de gangbare garantieperiode in Europa) en dit
voor installaties die te vinden zijn bij een gemiddeld gezin.
Een robot, die gebruik maakt van twee touwen die bevestigd zijn aan de hoeken van het dak, werd
onderzocht. Deze aandrijfmethode werd virtueel gemodelleerd om de prestatie te controleren.
Daarna werd ook een proof of concept gebouwd.
Het resultaat is dat de robot een veelbelovend alternatief is om toe te passen op kleine opper-
vlakken. Het systeem is tegelijk zeer beweegbaar, flexibel toepasbaar en tegelijk kostenefficient.
Volgens een ruwe schatting kan de robot zich in twee jaar terugverdienen bij vervuilingsverliezen
van 15% voor een gemiddeld gezin. Voor grote oppervlakken is het zelfs sneller rendabel. Echter
aan de betrouwbaarheid moet nog gewerkt worden.
v
Contents
1 Introduction 1
2 Literature Study 3
2.1 Photovoltaic energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Losses in PV systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Residential use of solar panels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Existing cleaning robots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Methodology 15
3.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4 Drive principle selection 18

4.1 Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2 Choosing the motors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5 Modelling the robot 25

5.1 Modelling the static forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.2 Modelling the dynamic forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.3 Multi-domain model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.4 Dimensioning parts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
6 Validating the models 36

6.1 Test setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6.2 Comparing the models step responses . . . . . . . . . . . . . . . . . . . . . . . . 38
6.3 Checking the limits of the system . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.4 Testing the braking qualities of the motors? . . . . . . . . . . . . . . . . . . . . . . 42
6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
vi
CONTENTS vii
7 Implementation and Applicability 44

7.1 Limits of the system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
7.2 Power consumption and battery . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
7.3 Path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
7.4 Brush . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
7.5 Additional electronics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
7.6 Complete robot and price . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
7.7 Calculation of return on investment . . . . . . . . . . . . . . . . . . . . . . . . . . 49
7.8 Secondary and tertiary goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
7.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
8 Conclusion and recommendations 51

List of Figures
2.1 PV system power output under different weather conditions. . . . . . . . . . . . . . 4

2.2 Transmittance over time (measured from 01/03 to 28/04 2011). Please note that the
lines only connect the measuring points. . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Transmittance decrease due to dry residue after evaporation . . . . . . . . . . . . . 7
2.4 Examples of panel degradation due to (a) haze and (b)hotspots without bypass pro-
tection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.5 Influence of fresh fallen snow and total snow height on the energy yield. . . . . . . . 9
2.6 The solarbrush robot without its brush. . . . . . . . . . . . . . . . . . . . . . . . . 10
2.7 Ecoppia E4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.8 An example of a Wash Panel system. . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.9 A person cleaning PV panels manually with waterfed pole. . . . . . . . . . . . . . . 13
3.1 Flowchart of the methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.1 Conceptual representation of a design with a horizontal rail and a vertical brush. . . . 19
4.2 Conceptual representation of a design with a horizontal rail and a horizontal brush. . 19
4.3 Conceptual representation of designs using two or four tethers. . . . . . . . . . . . . 20
4.4 Torque and power characteristic of a brushed dc motor. . . . . . . . . . . . . . . . . 23
5.1 Vector representation of the forces on the motor controlling the left tether . . . . . . 25
5.2 Plot of the forces on the left line (mass=1kg,tilt-angle=35 deg) . . . . . . . . . . . . . 27
5.3 Plot of the critical speed (inclination=35 deg, m=1kg, width=5m treshold=0.15 m/s) . . 28
5.4 Plot of the possible acceleration on the robot (inclination=35 deg, m=1kg, ) . . . . . . 29
5.5 Simulink model of the physical model of robot. . . . . . . . . . . . . . . . . . . . . . 29
5.6 Simulink model of the physical model of robot. . . . . . . . . . . . . . . . . . . . . . 32
5.7 The control loop for both the virtual as the physical model. . . . . . . . . . . . . . . 33
5.8 Screenshot of the robust controller . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6.1 Pololu Dual MC33926 Motor Driver Shield. Note that in the figure the shield is con-
nected to a Arduino Uno board . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6.2 Test set-up of the the small scale physical model . . . . . . . . . . . . . . . . . . . 38
6.3 Step function applied to the Y input of the controller in the middle of the surface . . . 39
viii
LIST OF FIGURES ix
6.4 Step function applied to the Y input of the controller on the left side of the surface . . 40
6.5 Step function applied to both X and Y input of the controller . . . . . . . . . . . . . . 40
6.6 Current drawn by the right motor in the diagonal movement experiment. . . . . . . . 41
List of Tables
2.1 Estimated daily and monthly power output in kWh for a panel of 1kWp located in
Sint-Katelijne-Waver, Belgium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Pros and cons of Solarbrush system . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Pros and cons of Eccopia E4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Pros and cons of Washpanel system . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5 Pros and cons of a sprinkler system . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.6 Pros and cons of manual cleaning . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1 Success levels for the project with explanation. . . . . . . . . . . . . . . . . . . . . 16
4.1 Ranking table for the design selection process . . . . . . . . . . . . . . . . . . . . 21

4.2 Pros and cons of DC and Stepper motor. . . . . . . . . . . . . . . . . . . . . . . . 22
6.1 Specifications at 12V for of DC-motors used in the experiment . . . . . . . . . . . . 37

6.2 Stall distance when pulling robot straight up . . . . . . . . . . . . . . . . . . . . . . 42
7.1 Stall distance for different roof settings . . . . . . . . . . . . . . . . . . . . . . . . . 44

7.2 Example of a possible cleaning path focusing on the lower edges. . . . . . . . . . . 45
7.3 Estimation of the total price and weight of the robot . . . . . . . . . . . . . . . . . . 48
7.4 Checklist list of the met secondary goals . . . . . . . . . . . . . . . . . . . . . . . . 49
7.5 Checklist list of the met tertiary goals . . . . . . . . . . . . . . . . . . . . . . . . . 50
x
Chapter 1
Introduction
Growing interest in renewable energy has led the solar photovoltaic (PV) industry to expand notably
in the last decade. In the year 2013 a staggering 80 terawatt-hours of electricity [1] was produced in
the European Union hence covering 3 percent of the total electricity demand. Because Photovoltaic
energy is an accessible technology it has become a popular investment for companies as well as
for residential users. Consequently, this demand has stimulated the research for increasing the
overall output power of PV systems causing laboratories all over the globe to work hard on making
the technology both more efficient and cost effective.
Regardless of the effort of the industry to shorten the payback time, a preventable loss namely
soiling is often overlooked. As a result a layer of dirt piles up on top of the glass reducing its
transmittance and therefore decreasing the power output of the entire system. The rate at which
the power reduces over time is rather unpredictable as it depends on various environmental factors
such as the type of soil, agricultural activity, precipitation, wind, bird droppings, ...
Common order of magnitude for losses due to soiling can be estimated between 3-8% but can go
up to 30-40% in dry and sandy climates. Seeing that precipitation plays a considerable role in the
cleaning capability it must be said that rainfall often does not suffice because some types of soil
cement and stick. The same counts for bird droppings which dont flush away either.
However cleaning solar panels is not always as straightforward. To begin with, there is the issue
of accessibility. Due to the fact that PV panels often are situated on dangerous and difficult to
reach places, it might be hard to clean them manually and it takes time to do it safely. Secondly,
cleaning a panel only once a year might not have a significant impact on the yearly energy yield
for the simple reason that dirt stacks up again in a short period of time making the difference
negligible. Especially if you need to contract someone to clean the panel for you, it might just not
be economical. However, leaving panels uncleaned might not be wise either since soiling can lead
to permanent damage of the glass limiting the lifespan of the installation.
Lastly, because cleaning the glass only seldom pays off and plants might be unreachable to clean,
the logical solution is to clean them automatically and autonomously. Sadly, there isnt an abun-
dance of such products available and therefore often built to purpose robots do the job. As a result
cleaning robots are only cost-effective on vast plants.
Therefore this paper is aiming at working out a concept design for an autonomous cleaning robot
that is both flexible as it is cost-effective in order to be economical at smaller scales. The way to do
this is by focussing on the most costly features of installing a robot to an array of panels.
The biggest factor that makes up the cost of a cleaning robot is the drive system. Cleaning robots
1
1 Introduction 2
are most commonly installed on rails along the panels. These make the installation relatively ex-
pensive for two reason. The first reason is the obvious fact that more material costs more money.
All parts are machined and often are built to purpose, in order to fit a specific plant. The second
reason is the labour costs for installing the track system to the installation.
Apart from rail based robots, there are other solutions for cleaning panels like sprinkler systems
or robots driving on caterpillar tracks made from suction cups. However, these solutions lack both
flexibility and cost effectiveness to be applied on a smaller scale.
The logical next step is to look for a more adaptable and economical drive principle for the robot
to translate itself over the panels. A concept then has to be investigated on the different design
criteria that come along with building such a robot.
The robot has to be able to handle all the forces involved with moving it and this for the different
angles at which solar panels can be installed. The driving principle has to be modelled in order
to make sure the demanded criteria are met and to simulate its behaviour on a commonly sized
installation. Later on, a physical proof of concept is made in order to validate the model and to
grasp of how it would work in real life.
The last part of the paper focusses on applying the drive principle to a concept that shows how it
could work in real life and to point out the strengths and weaknesses of the system. Also a rough
idea of the price and returns are given.
The research of the thesis was done during an abroad semester at Technische Hochschule Nurn-
berg and executed at their 3D visualisation centre.
Chapter 2
Literature Study
In this chapter a brief theoretical background will be given in order to define the problem. First
the focus will lay on the working of the photovoltaic cell. Next this chapter will go deeper into the
mechanisms of soiling and degradation as it is important to know what has to be done for cleaning
purposes. Lastly there will be a brief comparison of existing cleaning robots pointing out their
weaknesses and strengths.
2.1 Photovoltaic energy
Ever since the invention of the silicon photovoltaic cell in 1954 [2] it has played a big roll in the de-
velopment of standalone technology. Tapping into the inexhaustible energy source that is the sun
made it possible for the first satellites to be autonomous for a virtually infinite time. Nevertheless,
it took until the mid 70s before scientists started to recognize photovoltaic energy as a good can-
didate for future non fossil energy supply. Research has led to lowering the development cost and
system component cost but modern day research has also proven that there is still a lot of room for
improving the technology in terms of efficiency, size and manufacturing.
2.1.1 Working principle
When a photovoltaic cell is exposed to sunlight it absorbs the photons hitting the semiconducting
materials. Electrons are excited and move up to a higher molecular or atomic orbital. To dissipate
the extra energy, the electron can either go back to its original orbital, converting the excess energy
into heat, or it can travel through the material to an electrode, thereby cancelling the potential.
Regardless of the size, a cell will generate roughly 0.45 volts DC. This implies that the available
power generated by the cell will be strictly dependent on the area of the cell that is irradiated by
the sun and the material used for the absorption of the photons. To reach higher voltages, cells are
installed in series. The types of cells that are commercially available today can be divided into 3
main groups [2]:
Crystalline silicon cells: by far the most used bulk material for PV cells.
Thin Film cells: heavier but have a smaller ecological footprint.
Multijunction cells: still experimental, originally only used in space but terrestrial solar con-
centrators make them effective on earth now as well.
3
These three types can all be divided into different subgroups, using different materials and methods
dependent on the spectrum available and the cost-effectiveness of the cells. The two most used
materials are monocrystalline and polycrystalline silicon cells. The first has a higher efficiency but
is also more expensive compared to poly cells. However, due to technological improvement, the
price of mono cells has decreased and its application has grown again.
The output power of a cell is the product of the solar intensity (I 100mW /cm2 ) and the conver-
sion efficiency of the cell being around 16% for crystalline silicon cells. In 2014, scientists at the
University of Cambridge [3] succeeded at making solar cells based on a hybrid material that has
an efficiency of 95% proving there is still a serious improvement potential.
To protect the cells against the wear and tear of the operational environment, it gets encapsulated
with a material (most often glass or polycarbonate) that has both good optical qualities in terms of
transmittance and provide protection from impact, humidity, etc. Lastly they get hermetically closed
off with silicone.
2.1.2 Weather influences
The output power of photovoltaic cells is closely related with the weather, date and time [4] as
shown in figure 2.1.
Figure 2.1: PV system power output under different weather conditions.
The many parameters make creating a reliable model for output power a complex task. Before
installing solar panels, it is always good to consult a model to know how much energy a panel will
yield. Such a calculation model can be found on the website of the Institue for Energy and Transport
(IET) of the Joint Research Center of the European Commission. Their Photovoltaic Geographical
Information System (http://re.jrc.ec.europa.eu/pvgis/) makes it possible to generate a power output
estimation based on geographical location and mounting configuration. Table 2.1 is an example of
an monthly estimation done for Sint-Katelijne-Waver,Belgium.
Table 2.1: Estimated daily and monthly power output in kWh for a panel of 1kWp located in Sint-Katelijne-
Waver, Belgium
Ed: Average daily electricity production from the given system (kWh)
Em: Average monthly electricity production from the given system (kWh)
Hd: Average daily sum of global irradiation per square meter received by the modules of the
given system (kWh/m2)
Hm: Average sum of global irradiation per square meter received by the modules of the given
system (kWh/m2)
2.2 Losses in PV systems
As sunlight gets converted in usable electricity there are different losses lowering the systems
output power. To quantify the performance of a PV system, a performance ratio [5] is introduced.
This ratio represents the final yield (YF ) divided by the reference Yield (YR ). In essence it compares
the AC output power with the DC power coming out of the module, thus calculating the losses due
to the conversion. These losses can be linked to:
panel degradation(deg )
temperature (tem )
soiling (soil )
internal network (net )
inverter (inv )
transformer (tran )
system availability and grid connection network ( ppc )

Therefore, the performance ratio (PR) can be expressed as following:
YF
PR = = deg .tem .soil .net .inv .tran . ppc (2.1)
YR
The performance of inverters, transformer,... are increasing due to technological improvements.
However soiling (and in a smaller quantity degeneration) can be improved regardless of the tech-
nology installed, simply by keeping the glass clean. The next sections will go more into detail of the
mechanisms behind soiling and degradation
2.2.1 Soiling
When time passes, operational PV panels get covered with a thin layer of dirt, hence decreasing
the amount of light hitting the cells. The amount of power lost due to soiling is dependent on several
factors.
First of all it depends on the type of dirt deposited on the panel. Possible polluters [6] are sea salt,
pollen or matter from air pollution, agricultural activity, construction and other natural sources.
Secondly there is the influence of precipitation. Both the amount of rainfall and the time between
rain events change the layer pile up. A large amount of the accumulated pollution is washed away
by rainfall while in periods of drought losses increase rapidly. All these factors can be brought back
to the geographical location and the climate the system operates in.
A last important factor is the angle at which the panel is tilted. The lower the angle, the faster dirt
accumulates having a maximal effect when panels are installed horizontally.
Soiling decreases the transmittance of the glass which consequently affects the short circuit current
leading to average power losses that vary greatly:
In drier places like the middle east, the south of Europe, the southwest of America,... the layer
accumulation goes fast. In California, [6] the performance of some of the plants decreased
6.9% in a timespan of only 8 weeks. Because the soiling happens so fast, cleaning the
panels only once a year halfway through the summer drought period on average only yields
0.81% more annual energy while cleaning the sites regularly averaged a 9.8% increase in
annual energy. They also noticed that panels installed under an angle of less then 5 degrees
had up to five times the amount of soil compared to more heavily tilted panels.
According to another study done in Navarra, Spain [7] increased soiling on flat panels can be
dedicated to the fact that rain doenst flush away soil that effectively. Their experiments were
also executed in a dry climate and they noticed a performance loss of 8-22% in dry periods
due to soiling, with losses increasing about 0.1-0.2% a day. Rainfalls above 4-5mm seemed
to make a noticeable difference.
Similar experiments on Crete, Greece [5] found that soiling losses were 4-5% during winter
and 6-7% during the summer period, resulting in annual losses of 5.86%. There the longterm
effects of soiling are practically linear, the accumulation of deposited material can lead to
losses up to 30-40% when the panels are not cleaned at all.
For a central European climate there is a constant loss of 3%-4% according to experiments
done in Leuven, Belgium [8]. Their research focused on the transmittance of glass and they
noticed that in as little as 5 weeks the glass soiling reaches a saturation point as seen in
Figure 2.2. The experiments also proved that the saturation is caused by rain as they didnt
see the phenomenon on panels that were protected from rain by a roof.
Figure 2.2: Transmittance over time (measured from 01/03 to 28/04 2011). Please note that the lines only
connect the measuring points.
They also investigated the cleaning aspects on the transmittance of the glass. When glass
gets cleaned and the water is left to evaporate slowly, a dry residue is left behind together
with a layer of particles captured by the moisture (Figure 2.3). This could possible explain
why long-term soiling still increasing even with a lot of rain. It implies that using sprinkler
systems on non-hydrophobic glass might actually not be effective way of cleaning.
Figure 2.3: Transmittance decrease due to dry residue after evaporation

Traffic pollution causes particles with a small size (1-50 microns) to deposit on the glass and
stick easily when there is no rain. This effect will be aggravated by high humidity causing
a cementing effect on the surface. Even a small amount of such fine particles can reduce
transmittance by as much as 11%. [9]
It is clear that soiling is a very complex process due to the large variety of parameters. Conse-
quently, only rough estimates can be made for a given location.
2.2.2 Degeneration
Over time, PV cells lose some of their efficiency due to a various amount of non reversible damage.
A couple of reasons for this phenomenon are oxidation, delamination, cell cracks, irreversible soiling
and hotspots [10].
This last one is a phenomenon caused by shadows or localized dirt. Because the cells dont supply
energy they will start to act like a dissipater and heat up, leaving burned areas and damaging the
panels notably.
Irreversible soiling typically takes place at the lower edge of the panel. The polluted environment in
combination with water and the alkalis in the glass result in a loss of optical transmittance (haze)
along the lower edge of the PV module like in figure 2.4.
(a) (b)
Figure 2.4: Examples of panel degradation due to (a) haze and (b)hotspots without bypass protection.
Both phenomena cause a reduced short-circuit current and result in a reversed polarity, consuming
energy and increasing temperatures leading failure of the strings. To minimize the impact on the
performance, modern solar panels have a system with bypass diodes, shutting of cells in order to
protect the cells and the entire panel. As a consequence the bypassed zone will not be providing
energy. Both phenomena can not be totally ruled out but keeping the panels clean can certainly
help slowing down the degeneration of the modules.
2.2.3 Snow
Although the summer is the most advantageous period for solar panels, the higher temperatures
reduce the overall efficiency of the panels. Winter has a smaller share in the annual yield but the
effects of snowfall can clearly be seen in the production statistics. Snow can be seen as an extreme
form of soiling. Fresh snow stacking up limits the amount of radiation penetrating through to the
cells, reducing the final yield. Even when the area is only partially covered with snow, the impact is
still substantial due to the bypass diodes disabling cells. On tilted panels the snow starts slipping
and gradually the output power increases again, but on flat panels the snow will cover the panel
until it melts again.
Researchers in Munich [11] analyzed data captured from a local 38100 m2 big PV array in order to
find the effect of snow on the annual yield of the plant. According to their results (see figure 2.5) the
impact depends on various parameters including the height of snow, its weight and the tilt angle of
the modules.
Figure 2.5: Influence of fresh fallen snow and total snow height on the energy yield.
They concluded that the annual yield loss due to snow events varied between 0.3-2.7%.
Clean panels can at the same time profit from snowy situations. Snow reflects solar radiation from
the surroundings onto the PV array increasing the amount of harvestable energy. This means that
the potential loss of snowed panels might even be larger.
2.3 Residential use of solar panels
For calculation purposes it is helpful to have some estimated values regarding residential electricity
consumption. On the website of Vreg [12] (Flemisch regulator of electricity and gas) the following
data can be found:
An average Flemisch family consumes around 3500 kWh each year.

The average unity electricity price for an average consumer in Flanders in 2014 was e0.18/kWh
Consequently, the average family spends e630/year
Considering a family tries to produce as much energy as it consumes, it is possible to calculate the
surface of the PV array needed:
3500kW h
30m2 (2.2)
120kW h/m2
According to the Flemisch government [13] the the installation cost for 1kWp is on average 1700
euro (6% VAT included). As we know from figure 2.1 1kWp delivers 960kWh/year which means that
a system generating 3500kWh would cost roughly 6600 Euro. Given the current electricity price
the investment should pay itself back in around 11 years.
2.4 Existing cleaning robots
The cleaning of PV arrays is mostly done on large scale with big and heavy, custom made machin-
ery. They generally focus on cleaning as many square meters in the least amount of time and often
require an operator. The following list focuses on robots that are relatively smaller and dont need
an operator.
2.4.1 Solarbrush
The Solarbrush [14] is a light-weight, autonomous robot for dry cleaning solar panels or glass in
dry environments. Using tracks made o suction cups, it moves over the panels making it possible to
cross gaps up to 30mm and working on surfaces tilted up to 35 degrees. Different kind of brushes
can be attached to the front of the robot in order to brush away dust, sand and dirt. Power comes
from a rechargeable battery.
Figure 2.6: The solarbrush robot without its brush.
Pro Con
Small Needs operator
Highly autonomous Requires safety rope on tilted panels
Expensive suction cup system
Batteries need to be reloaded
Table 2.2: Pros and cons of Solarbrush system
2.4.2 Ecoppia E4
The Ecoppia E4 [15] is a fully autonomous robot that uses microfiber brushes to sweep the dust off
of panels. Its designed for large rows of panels situated in dry and sandy environments.By making
use of gravity the brushes move downwards spinning, creating an airflow that helps blowing off the
dust.
Figure 2.7: Ecoppia E4
The robot makes use of an on-board solar panel and battery to store energy that allows the robot
for cleaning at night. It also recovers energy when the robot descends. Another interesting feature
is the positioning of the solar panel on the bottom of the rail, allowing it to be cleaned by the brushes
as well. Both horizontal and vertical translation happens using a guide rail and wheels.
Pro Con
Energy efficient Only for large arrays
Highly autonomous Lot of construction needed
Table 2.3: Pros and cons of Eccopia E4
2.4.3 Washpanel systems
An Italian company called Wash Panel [16] produces robots that clean arrays of PV panels by
moving a vertical brush horizontally over a row of panels. Having a length of 1 to 16 meters and
containing a 12V battery it can be deployed automatically. Also a water hose is attached for wetting
the panels while cleaning. However the robot rather doesnt need a track like most panels have.
Pro Con
Very robust Only 1 row of panels
Highly autonomous Custom made
Table 2.4: Pros and cons of Washpanel system

Figure 2.8: An example of a Wash Panel system.
2.4.4 Sprinklers
Sprinkler systems are often used in dry areas to keep panels clean. It has the same cleaning effect
as rainfall and will clean panels at a relative low cost. Like most systems, the one found on the
website of Heliotex [17], exist out of a water filtration system and soap dispensing system. They
also mention that water hardness has to be tested to and optionally a water treatment is advised
for locations with hard water.
Although this is a relatively good working system, there are a couple of drawbacks to it. First of,
there is the large amount of water it uses as it cleanes the panels multiple times a day dependant
on the scheduling. This is needed because as the panels get dirtier it becomes trickier for the
sprinklers to remove all the soiling. Secondly both filters and soap level have to monitored and
taken care of both taking time and money. Lastly there is the fact that wet panels make dirt stick to
the panels [8] causing the panels to get dirtier in the long run. This system is suitable for very dry
sandy areas where sand stacks up very fast but will not clean as thorough as a brush system.
Pro Con
Small Needs maintenance
Highly autonomous Uses a lot of water
Little wear over time Might only be effective in dry sandy places
Table 2.5: Pros and cons of a sprinkler system
2.4.5 Manual cleaning
Technically this is not a robot but as this is the one of the most used ways of cleaning residential
solar panels today it therefore important to mention as well. Using a so-called waterfed telescopic
pole and a special non scratching brush, the panels gets cleaned manually. Standing on ladders or
climbing on roofs, this is often a job for a professionals as it is risky and requires the right material.
However as mentioned in section 2.2.1 this method might have little or no influence on the yearly
yield unless it gets done regularly. Especially if the job is done by a contractor this might be very
bad investment there the labour costs will exceed the profitable return.
Pro Con
Thorough labour intensive expensive
equipment needed
only beneficial if done regularly
dangerous
Table 2.6: Pros and cons of manual cleaning
Figure 2.9: A person cleaning PV panels manually with waterfed pole.
2.5 Conclusion
Although, at first glance losses caused by soiling dont seem to have a huge impact on the
cells efficiency, they are easy to minimize by cleaning panels at an economically optimized
frequency.
Predicting the exact profit of cleaning the panels regularly is as of today still difficult. It
depends on many factors of which precipitation and the presence of pollutants are the most
important. This can only be estimated using statistical data.
Existing robots focus mainly on large arrays and are in general heavy and unsuitable for
installing on smaller arrays i.e. residential roofs. Only a sprinkler system has the ability to
operate profitable at a residential scale but this only in sandy dry areas. In areas like these it
might not always be tolerable to use this amount of water.
But most importantly it are the smaller scale setups in central European-like climates that
dont have a solution for cleaning their PV system efficiently. Losses might not be as extreme
as in warm dry climates but under the wrong circumstances a cleaning device can be justi-
fiable too. Traffic congested cities like Antwerp and Brussels that have a lot of fine-particle
pollution will see their panels get dirtier faster. Another example might be panels in an agri-
cultural environment causing a lot of dust to settle on the glass. These panels could have
losses up to around 10% a year and more if they would otherwise not be cleaned at all.
Therefore investigating the possibility to develop an autonomous cleaner optimized for small arrays
is interesting. The biggest designing factors should be size, flexibility but most importantly cost
effectiveness.
Chapter 3
Methodology
The chapter gives in the first place an insight to the used methodology of the project. Also, the
goals for the project are set and a strategy is outlined.
3.1 Methodology
The main reason for cleaning photovoltaic panels is increasing the profitability of the investment.
This means that the robot needs to have a reasonable payback time and it should give a satisfying
return-on-investment as well. Thats why the first and most important step in the project was to find
an element that could improve cleaning robots in terms of both flexibility and profitability.
One of the biggest costs of the system is the drive system and the installation costs. That is
why firstly an optimal drive system was sought. (chapter 4) Choosing motors, gears, spools and
feedback system define not only how the robot can move over the roof, it also defines the different
possibilities of cleaning the roof. For example, in order to brush soil of the roof in dry climates, the
robot has to be able to drive in straight lines and reach all places with a straight brush while a less
controllable robot would need other solutions.
In order to simulate and dimension the parts, multiple models were made (see chapter 6) in order to
investigate the chosen drive principle further. The models both investigate the static as the dynamic
behaviour of the system. The models can be used to built and test an appropriate driver to control
the lines and creating a path.
After modelling the system and dimensioning the parts, a test set-up was built in order to validate
the models and check the controllers performance. When the models turns out to be reliable, the
limits in both movement as controllability can be assessed and conclusions regarding the applica-
bility of this drive principle can be made by studying the computer simulations.
Next a cleaning strategy is forged based on the knowledge about soiling and the findings on the
movability. Components are laid out and an appropriate brush is sought. Both dimensions as
functions will be influenced by previous steps and if necessary there has to be done some iteration
in order to find an optimal combination.
When this step is done all the components should be known and it will is possible to make a
reflection on the feasibility, the return on investment and price.
15
3 Methodology 16
Figure 3.1: Flowchart of the methodology
3.2 Goals
Before starting the design phase, specific goals have to be set in order to define the concept. The
most important factor in this project is undoubtedly the payback time and will be the main objective.
This can be broken up in two parts, one being the cost of the device and its installation, the other
being the energy won and therefore the returned money.
A justifiable payback time could be 2 years, there European electronics manufacturers are obliged
to provide 2 years of warranty. In other words, return on investment would be guaranteed.
Ideally the device should work for an average family sized system with a low amount of losses
(3%-4% a year). However, this might be raising the bar too high. By defining different situations
and their goals, the success of the final product can be quantified.
Level Size Loss Explanation

1 30m2 3%-4% Average family with average power loss
2 30m2 10%-15% Average family with worst case losses
3 60m2 10%-15% Small business with worst case losses
4 60m2 20%-30% Small installation in a southern climate (e.g. California, Saudi Arabia,...)
Table 3.1: Success levels for the project with explanation.
Each level in table 3.1 represents a high loss in energy and consequently loss in money. This
means that for every higher level the robot will pay itself back quicker and the lower the success
level, the more applicable the robot is.
Apart from the cost there are other designing criteria that have to be fulfilled to be a useful (con-
sumer) product:
Ability to clear gaps between panels
Should work on ramped roofs
Shouldnt damage the panels
Has to be weather resistant
Should be as small as possible
Has to be as autonomous as possible
Has to be flexible
3 Methodology 17
Other tertiary goals can be added. They dont have to be met but they could add extra value to the
product:
Ability to clean flat panels
Ability to move snow
Docking station
Interchangeable brush so it can be used in various climates

Chapter 4
Drive principle selection
In this chapter different design options for the drive principle are put side to side. The drive principle
will define which components are needed. Also the flexibility and adaptability are dependant on this
design.
4.1 Mechanics
Selecting a drive principle starts with the moving dynamics. Different concepts are compared and
then rated on their performance regarding the set goals.
4.1.1 Rail
Horizontal Rail with Vertical Brush
A vertically aligned brush is mounted on a structure that can move horizontally over a rail system
(See figure 4.1). Because the path of travel is defined, the brush can only clean one row of panels,
which is for residential arrays often not enough. The brush could be angled a bit so dust gets swept
towards the bottom and not just transported sideways.
Pro: Contra:
-Few parts needed -Rails might have to be installed on the
roof
-Winterproof -Needs a large brush
-Stable -Can only clean 1 row of panels
-Cleans both flat as tilted surfaces -Cant really brush off dust
-Clears Gaps
Horizontal Rail with Horizontal Brush
A possible variation on the horizontal rail design could be obtained by replacing the vertical brush
by an axis on which a horizontally aligned brush could move (See figure 4.2). Although this makes
the design more complicated it can sweep of dust more easily by just brushing it downwards and
off the panel.
18
Pro: Contra:
-Smaller Brush needed -Rails might have to be installed on the
roof
-Winterproof -Can only clean 1 row of panels
-Stable -Cleaning takes longer
-Can sweep dust off
-Clears Gaps
Figure 4.1: Conceptual representation of a design with a horizontal rail and a vertical brush.
Figure 4.2: Conceptual representation of a design with a horizontal rail and a horizontal brush.
4.1.2 Tethered
Two Parallel Tethers
The robot hangs with two parallel tethers to a rail that allows the robot to move sideways. (See figure
4.3a) The design allows the robot to adapt to the panel formation and to cover multiple coplanar
rows of panels. The design would be very effective for arrays installed on tilted roofs. A drawback
of this gravity requiring method is that it wont work if the tilt angle is not sufficient enough.
Pro: Contra:
-Clears gaps -Tilt angle needs to be sufficient
-Can clean large areas -Still needs a rail
-Relatively easy movement -May need extra sensor keeping it from
deploying in bad weather.
-Both the rail carriage as the cleaning
device need power and need to coop-
erate.
Two Corner Tethers
Instead of mounting the tethers to a horizontal rail they can be attached to the upper two corners
of the area that has to be cleaned. (See figure 4.3b) This means it doesnt need a rail-carriage
anymore decreasing the amount of parts. The robot would need more electronics to monitor the
translation and also the travel path is complexer to define.
Pro: Contra:
-No need of rail -More complex translation
-Flexible -Tilt angle needs to be sufficient
-Needs less motors -Needs to monitor weather
-Easy to install
Figure 4.3: Conceptual representation of designs using two or four tethers.
Four Corner Tethers
The robot could also use four tethers attached to the corners of the area that has to be cleaned.
(See figure 4.3c This would mean it could move regardless of the tilt angle. This makes the design
more flexible but at the same time more complex. Another big drawback is that no matter where the
robot positions itself, there is at least one tether running over the panels casting a shadow, possibly
leading to losses.
Pro: Contra:
-Flexible -At least one cable stays on panel while
in rest casting a shadow
-Easy to install -Many motors
-Complex
4.1.3 Untethered
The robot wouldnt necessarily be attached to the panels, it could be possible to use wheels for its
translation. An untethered design would only be usable if the panels are mounted flat or under an
only small angle because it would otherwise need a safety tether keeping it from falling of the roof
and possibly causing damage or even injury.
Pro: Contra:
-Very mobile -Can only clear small gaps
-Flexible -Possibly still needs safety tether
-Needs that have grip on the panels
without damaging them
4.1.4 Selection Process
For the selection of the best design, a weighted scoring system was applied taking several factors
into account:
Flexibility: Is the the design usable for different panel configurations and how well is it appli-
cable on the average households roof.
Complexity: How complex is the machine or the parts on which it has to rely.
Cost: An intuitive guess of the relative price.
Size: How big is the the machine and its needed construction.
Cleaning area: What area can the machine clean relative to its own size.
For each criteria the different designs got ranked from 1-6 where 6 is the best score.
Table 4.1: Ranking table for the design selection process

The drive principle with two corner tethers scores very good on every criterion. Therefore it is
interesting to investigate this drive principle further and find out whether it really is a possibly more
preferable drive mechanism or not.
4.2 Choosing the motors
The key element in the two-tethered drive principle are the motors. The weight of the entire robot
hangs on the two ropes and the forces get transferred by a spool into the motors as torque.
For the purpose of building a small stand-alone device there are two possibilities regarding motor
choice, each having their pros and cons. (See table 4.2)
Table 4.2: Pros and cons of DC and Stepper motor.
A stepper motor would have a couple of advantages in terms of controllability over a DC motor
there each step has a defined angular rotation. Also, because it could deliver a lot of torque at low
speeds and having reasonable currents at stall torque would make it a reasonable option.
However, for making the robot be a product that is easy to install, the motors are preferably built into
the robot. More so, because if the motors would be positioned on the edges of the roof, the robot
would have to communicate with them. By consequence they have to carry their own weight, a
high power to weight ratio is therefore wanted. In addition, the robot has to be provided with energy
as well. Therefore a battery would need to be installed in the robot. The power consumption then
would have an influence on the battery capacity and therefore also on the weight and costs of the
entire system. This is what makes a DC-motor a more attractive candidate for for building it into
to the robot. The drawback of this system though is that driving a DC-motor accurately is more
challenging as driving a stepper motor.
However, this paper is focused on the use of DC motors for the following two reasons.
It is interesting to investigate the controllability of the DC motors. Stepper motors are rela-
tively easier to control. This has already been demonstrated in hanging plotters.
Both the price and weight of DC motors are preferable over stepper motors.
4.2.1 Characteristics of the DC motor
In order to do calculations regarding the motors it is important two know some key characteristics
of the motor. The performance of a permanent magnet DC motor is defined by two characteristic
parameters, namely stall torque and the no load speed. These two characteristics for a given
voltage define a straight curved torque-speed characteristic. As a result, the speed at which the
motor has its maximum power is exactly at half the no load speed and half the stall torque. (See
figure 4.4 [18])
So when applying a DC motor it would be beneficial to use it at around half the maximum power
speed. This can be obtained by choosing an optimal spool size for the robot. A bigger spool would
reduce the motor speed and increase torque and vice versa. This way the speed can be tuned
optimally regarding to the motors RPM.
Figure 4.4: Torque and power characteristic of a brushed dc motor.
In general, the no load speeds of small DC motors are much too high for this purpose. However,
motors with built in gearboxes are more appropriate, both lowering the speed and increasing the
output torque.
4.2.2 Feedback system
There a DC motor does not have a feedback system, another way of retrieving information about
the motor speed is needed. An easy way of doing this is, is by attaching encoders to the end of the
motors. By doing this the current speed of the motor can be measured. By integrating this signal
over time, the position of the device can be calculated.
A drawback of using encoders is that the robot would have to be programmed and calibrated in
order to calculate its current position. This because the current position is dependent on the length
of both ropes and the width between the attachments to the roof. There the encoders can only read
the speeds of the motors, the device can only track difference in position and consequently has to
know were it is located when it starts.
There the essence of a small robot would be for it to be a flexible consumer product, it might be bet-
ter to find another feedback system. A good example for this would be an accelerometer. However
this system is not always reliable there it depends on the double integration of the accelerometer
data. This means the positional data can drift off very fast. Although this system would make a
better candidate, it will not be used further on in the experiments because the encoders are more
reliable for test purposes. However, the encoders could in the end be replaced by an accelerometer.
More about this subject is found in section 7.5.2.
4.3 Conclusion
There are a lot of options for moving around the robot. Many of them require custom made con-
structions for attaching it to the roof and also the robots often have to be made to purpose. This is
why tethered drive principles score very well in the selection process. It provides a lot of flexibility
while being easy to install.
However this drive principle has a couple of drawbacks. First of all, when the robot goes higher, the
torque the motors have to deliver increases. This means that the lines have to be attached higher
than the highest point it wants to clean. Secondly it has to be designed in such a way that the
movement is well controllable and reliable. This means it needs to have a reliable feedback loop in
order to provide good enough information to keep the robot travelling on the right path.
In order to investigate if this drive principle could work in practice and the forces on the motor dont
become too high, a computer model has to be made for the purpose of dimensioning the motors
and to test different ways of controlling the robot. Lastly there is the challenge of designing a
controller that is both flexible and reliable.
Chapter 5
Modelling the robot
In order to select motors that suffice for moving the robot around, a mathematical model was
designed, that can process the different variables like roof angle, weight, etc. and can output the
systems performance. Later also a multi-domain model is developed.
5.1 Modelling the static forces
Figure 5.1: Vector representation of the forces on the motor controlling the left tether
First of all is it interesting to know what the forces on the lines are at any given place while the robot
is stationary. This model gives a feel for the systems performance. The data can also be used to
validat if the motors can deliver the stationary torque.
The model has two input variables:
Mass: the heavier the robot becomes the higher the forces will be
Tilt-angle: the inclination of the solar panels changes the forces on the lines.
To turn the problem into a 2D problem the massvector can splitted into its two components. The
vector perpendicular to the panel can can be ignored if the static friction is neglected, the vector
25
parallel to the panel can be described as:
m = cos(90 tilt) mass (5.1)
Now it is possible to describe all the forces in a 2D plane. For the purpose of making the calculations
easier the origin is placed at the upper attachment point of the left line and the y axis pointing down.
(See figure 5.1)
The forces can be describes as following

FAy + FBy = m

F = FBx
FAx

Ax X
= (5.2)
FAy Y
FBx w X

=
FBy Y

This can be solved:
XFAy = Y FAx (5.3)
X
(m FBY ) = FBx (5.4)
Y

(m FBy ) Y wX
= (5.5)
FBy X Y
m
FBy = (5.6)
Y wX
1+
X Y

wX
FBx = FBy (5.7)
Y
FAy = m FBx (5.8)
Total forces on the ropes are then: q

FA = 2 + F2
FAx (5.9)
Ay
q
FB = 2 + F2
FBx (5.10)
By
Because the plots of both lines are symmetrical, only the left one is generated. Also, because
the forces are determined by the angle at which the lines are pulling, the length and width can be
normalised.
Out of the plot (figure 5.2) can be concluded that the higher the robot has to move the forces
become higher too. For the gradual scale of the plot a threshold of 15N was used. In reality the
forces theoretically become infinitely high at the top of the roof.
Figure 5.2: Plot of the forces on the left line (mass=1kg,tilt-angle=35 deg)
Another conclusion that can be made is that, to practically apply this system, the lines will have to
be attached higher than the panels themselves due to the high forces on the top edge of the plant.
5.2 Modelling the dynamic forces
Because the forces on the robot vary at every point in the plane it is necessary to calculate the
speed at every point. Also there has to be taken into account that the speed of the motors depends
on the load. Furthermore the robot has to be able to accelerate in every given direction in order to
change its trajectory. In other words, at a minimum required speed, the motors should have excess
torque.
The model needs the following input variables.
Mass: the heavier the robot becomes the higher the forces will be
Tilt-angle: the inclination of the solar panels changes the forces on the lines.
Spool radius: the radius of the winding spool will affect the torque affecting on the motors
RPM: Rotations per minute the motor turns without load.
Torque: Maximum torque the motor can handle at 0 rpm.
There both speed as excess torque have to be looked after, the model is divided in two parts, one
calculating the speeds, the other one checking the possibility to accelerate.
5.2.1 Calculating the work area based on critical speed
Because the lines of the robot are angled, the maximum speed of the robot varies at every position.
To make sure the robot can move at a reasonable speed at all times, a critical speed graph can be
plotted. This speed plot takes a maximum motor rotation speed as input and then calculates for
each point what the critical minimum speed is.
To calculate the minimum speed at each position the script takes a maximum motor speed as input
and creates a parallelogram of the two speed vectors. The direction of these vectors is defined by
the lines angles. Then the script checks in which direction the smallest resulting vector is situated
and takes this amplitude as the critical speed at that point. After the plot is made, a threshold speed
can be applied to define the working limits of the system. To give an easy overview over the plot,
the width of the roof is normalised.
Figure 5.3: Plot of the critical speed (inclination=35 deg, m=1kg, width=5m treshold=0.15 m/s)
It has to be remarked that this part of the script doesnt check whether the motor can deliver enough
torque to reach this speed. Therefore the script has to be expanded to check the available torque
at this minimum speed. This will be done in 5.2.2
5.2.2 Calculating the work area based on critical acceleration
Because in the calculation based on the minimum speed, torque isnt taken into account, it has
to be checked separately. To do this the script assumes the DC motor has a linear characteristic.
This way it calculates the torque available at the rotational speed given in the speed model (as in
figure 4.4). This way the available force is calculated. Then similar to the calculations in 5.2.1 a
parallelogram of the available force can be calculated. Then the gravitational vector of the robots
mass is subtracted from the vector parallelogram. This way a resulting parallelogram is created
that represents the excess force or in other words the available acceleration of the robot.
If the origin lies within this parallelogram, the robot can accelerate in every direction. However when
that is not the case, the robot is possibly not able to reach this speed or can not change its direction
in every direction. This information can then be plotted as seen in 5.6.
Figure 5.4: Plot of the possible acceleration on the robot (inclination=35 deg, m=1kg, )
The information of 5.3 and 5.6 can then be compared possibly plotted in one plot showing the
dynamic limits of the system.
5.3 Multi-domain model
The matlab model in section 5.2 can be used as a quick tool for iteratively finding a suitable motor
and calculating corresponding spool size. However, in order to test the controllability of the system
and the design of the driver, a multi-domain model was worked out.
Figure 5.5: Simulink model of the physical model of robot.

The multi-domain model was developed using Simulink and in particular its Simscape environment.
This makes it possible to integrate both the electrical and mechanical elements in one represen-
tative model. At the same time this environment can make a visual representation of the physical
elements making it easier to interpret the data.(Figure 5.5)
The real system would have 2 inputs (PMW controlling the input voltage of the 2 motors) and 2
feedback outputs (the speeds of the motors as seen by the encoders). The Simulink model uses
the same in- and outputs. This way the controller can be built around it as if it were the real system.
Henceforth the model was later validated within the Simulink environment by replacing the model
by a micro-controller. (see 6)
The Simscape environment also allows the user to use Matlab variables. This way a small Matlab
script was used as a set up file for initiating properties like size, weight, roof angle,...
Just like in the real world the machine consists of multiple parts. Each of them have their own
limitations and variables. Estimating the correctness of the results provided by this subsystems is
crucial for estimating the usability of the entire model.
5.3.1 The motors
The motor simulation block consists of two sub parts. The first part consists of all electrical com-
ponents (blue color in Simscape). They simulate a PWM, which controls the amount of voltage
delivered to the motor based on a reference voltage by quickly switching the power, and an H-
bridge, which allows to change the direction of the spinning motor by inverting the voltage. Lastly,
the current goes from the H-bridge to the motor which converts the electrical energy into mechani-
cal rotational energy.
Because all the lines in the model represent real electrical connections, it is possible to add both
voltage as current sensors to capture their values.
The second subpart (yellow) represents the mechanical component of the motor starting with the
rotational energy provided by the shaft of the DC-motor. Then a wheel and axle block is added for
simulating the conversion from rotational energy to translational energy by the spool. However, this
block simply translates the rotational speed into translational speed without giving feedback about
the torque. That is why an ideal torque source block is added as a feedback for the load.
Just as in the electrical circuit, a rotational sensor is added extra to give feedback about the rotation
of the motor.
An anomaly in the model is that the motors gearbox introduces a lot of friction were in the virtual
world the motor has no gearbox and consequently no friction. This results in a torque that is not
perfectly correct. However, the errors introduced in the model are within the limits of acceptability.
Nevertheless can this torque be simulated by adding a rotational friction block to the simulation.
However this made the calculations noticeably heavier without having a big difference on the results.
This friction might not have a big influence on the motors rotation when being powered. However,
when torque is applied on the output shaft in order to turn the engine, it feels way more friction due
to the amplification by the gears. This is effect is often called backdrivability and is not implemented
in the model.
Aside from the torque difference, the current turns out the be far off compared to the expected
characteristic. This is also caused by the gears resulting in another current-torque characteristic.
However this issue can be passed by calculating the current by dividing the measured torque and
the slope coefficient of the characteristic.

I= = (5.11)
k stall /(Istall I0 )
5.3.2 The mechanical model
For the mechanical part of the model a couple of solid bodies, each having a mass and dimension,
represent the different parts of the robot and the roof. To make them interact with each other they
are linked by joints of all sorts. In addition, coordinate transformations are added to move the
coordinate centres from the bodies center of gravity to pivoting points.
However, there is no block for simulating rope in the Simscape environment. Consequently a less
accurate substitute hqd to be introduced. That is why the rope is represented by a prismatic joint
providing the linear motion. Also, the joint feeds back the information about the torque to the ideal
torque source in the motor model.
The problem with this simplification is that a prismatic joint is a representation of a solid joint. In
other words, it can not give slack the way a cord does but instead pushes the robot down. This
in combination with the backdrivability problematic of the motor make that the controllability of the
robot and the model are different.
The biggest difference can be seen in the usage of the H-bridge. Because the model uses an ideal
motor with no backdrivability issues whatsoever, the H-bridge doesnt have to change the polarity
of the voltage in order to let the robot go down. Instead, by just lowering the amount of voltage
provided to the motor, the weight of the robot will pull hard enough to overcome the torque of the
motor and therefore the spool will unroll. Using the real motor however, the polarity of the current
provided to the motor has to be changed by the motor controller.
Furthermore is it not possible to simulate the friction between the roof and robot directly because the
simulation is in essence a 2D simulation and thus lacking the normal component of the gravitational
vector. This is why an extra planar joint was added to which a force can be applied. This force,
opposing to the direction of movement, is equal to the normal force times the kinetic friction factor.
The static friction is not taken account for. For the downwards movement of the robot it would be
important to know wetter gravity can pull the robot down or not but because the simulation doesnt
allow for downwards movement it would only make the simulation unnecessarily heavier.
However, because this part made the calculations too heavy, this part was disabled. As a replace-
ment a small amount of damping was added to the joint in order to filter out the swinging motion of
the robot, which normally would be done by friction.
Figure 5.6: Simulink model of the physical model of robot.

5.3.3 Control Loop
The control loop for the virtual model can later be reused for controlling the real model.
Just like in standard control theory the control loop consists out 4 parts:
Process: This is the real/virtual model and has two inputs (motor voltages) and two outputs
(rope length).
Measurement: The measured rope lengths are translated from polar to Cartesian coordi-
nates. To do this calculation, the script also has to know the distance between the two
attachment points, hence the third input to the measurement block.
Controller : The controller takes the two Cartesian coordinates as inputs and then allows
a pathplanner to provide an x and a y setpoint. It then compares both and calculates the
necessary adjustments.
Adjustment: A voltage level is send to the motor controller as and the motor speed is altered.
Figure 5.7: The control loop for both the virtual as the physical model.
Controller
The controller is the most important part of the steering. It takes the current Cartesian coordinates
as input compares it with the setpoint coordinates and then translates the information into an output
voltage for the pwm controller, thus driving the motors.
A big problem with the controller is that the system responds differently at the top of the roof than
it does on the bottom of the roof. This has as a result that, either the controller has to adapt to the
current situation, in other words adapt to the robots position on the roof, or it has to be so robust
that no matter were on the roof the robots is situated, it still gives a tolerable precision.
The first option, making it adapt is in reality very complicated to achieve. The controller would need
so much input information (roof width, response at different places,...) that it would need a lot of
programming in order to customise the robot to every roof.
Instead, a robust controller only needs the current coordinates and and a setpoint in order to calcu-
late the needed voltage and is therefore very flexible. However, because the robot uses encoders
on the motors to detect its position, the block converting the encoder information to a position needs
to know the distance between the two attachment points on the top of the roof. Other reference
systems might not need this information making this type of controller the best option.
Of course, the controller is only robust enough for a certain area, especially if the robot is applied to
a long roof with small space between the two attachment points, the controller will start to struggle.
This kind of form is not common for roofs and thus not relevant for this application.
The controller used for the model and the experiments (see figure 5.8) consists out of a couple
parts. First it starts by comparing the setpoints with the current positions and applies a PID con-
troller to this signal.
Figure 5.8: Screenshot of the robust controller
Then the controller compares both the x and y errors and decides what both motors have to do. If
the robot has to go up, both motors will get the command to increase their torque and vice versa
when they have to go down. When the robot has to go to the left, the left motor will get the command
to give more torque while the connection between both the two lines in the controller makes the
right motor take back some torque. This results in a robust way to control both output voltages.
Lastly, the controller has a saturation block in order to keep the controller from outputting a value
outside of the -12V/12V range needed for the motors.
This technique has one drawback, although it makes the robot to the setpoint as fast as possible, it
doesnt do it in a straight path. This means an extra pathplanner is needed to create a smooth path
between two wanted points in order to make the machine go straight.
Pathplanner
The pathplanner calculates, as the name suggests, the path the robot has to follow. For the experi-
ments, a very simple planner was used. Based on a given speed and a given begin and end point,
it determines where the robot has to be at every point in time.
Because in the experiments the robot only has to travel short periods of time, this planner suffices.
However in reality, it would be better to use a pathplanner that also takes the current coordinates
as inputs and recalculates the path every time so if the robot gets stuck for some reason the
motors wouldnt store so much energy in the ropes until the robots shoots loose and tries to move
uncontrollably fast to its current setpoint.
Also, experiments were performed with trying just simple coordinates as the next setpoint in stead of
calculating the whole path. However this approach didnt work well because of the unpredictability
of the path in between the begin and set point. This would in reality lead to parts of the roof not
being cleaned.
5.4 Dimensioning parts
In order to dimension the motors, different motors were tested iteratively. First of all a estimated
weight had to be assumed in order to do the calculations. For this purpose a weight of 1 kg was
assumed.
First, the mathematical models could be used to roughly estimate the needed torque and to find a
suitable RPM of the motor. The spoolsize, which is also a variable in the models, was also iteratively
altered in order to further fine-tune the rotational speed of the motor shaft.
By trying out different combinations the following estimations were made.
Estimated torque: 1nm
Estimated no-load motor speed: 200 400RPM
Spoolradius: 40mm
5.5 Conclusion
Different models concerning the simulation of the robot were made. The first two, mathematical
models, help with roughly dimensioning the parts needed for building the robot. They take into
account the necessary speeds, angles and roof dimensions in order to roughly state the limits of
the system with given motor characteristics.
Secondly a multi-domain model was created in order to simulate the real robot. It allows the user to
develop a control loop without having to test it on a real model, risking loosing control and breaking
it.
Next a controller was developed with the purpose of being as robust as possible. This way the robot
is flexible and easy to program.
If the models are correct, the also proof that it is possible to make drive principle based on ropes
and DC-motors. However, therefore the models have to be validated first by making a physical
model and comparing the results.
Chapter 6
Validating the models
This chapter describes the tests done on a physical model in order to validate the earlier made
models. By doing various small-scale tests both the controllability as its operational area were
analysed.
6.1 Test setup
6.1.1 Control
For controlling the robot an Arduino mega 2560 microcontroller board was used. This board is
based on the Atmel ATmega2560 microcontroller. With 54 digital input/output pins and 16 analogue
ones, the board has enough flexibility for this project. [19] Additionally, the board has 6 interrupt pins
(pins 2,3,18,19,20). These are especially important for the operation of the encoders counting the
motors rotation.
Everytime a pulse arrives from the motors encoder, an interrupt will is triggered, consequently
pausing other tasks so each pulse gets registered by the microcontroller. If not connected to an
interrupt pin, several counts per revolution of the motor would get lost and therefore the position of
the robot would be estimated falsely.
Another reason for taking the Arduino mega is its compatibility with Mathworks Simulink. In contrary
to the Arduino Uno, the Mega allows the user to interact with a Simulink model in real-time. By
run the model in so called external mode, all information can be send to and received from the
microcontroller in discrete time-steps.
A third advantage of the Arduino is that there is a vast amount of information and examples available
on the internet because it is often used for educational purposes.
6.1.2 Motors
The brushed DC-motors used in the experiments are 50:1 Metal Gearmotor 37Dx54L built by
Pololu. Weighing 218 grams each they closely approach the requirements found earlier in the
models for a robot with an estimated weight of 1kg.
The motors are equipped with quadrature encoders. This two-channel hall effect encoder mounted
on the shaft has a resolution of 64 counts per revolution. Because of the 50:1 ratio of the motor
36
Free-run speed 200 rpm

Free-run current 300 mA
Stall current 5000 mA
Stall torque 170 ozin (1.2nm)
Table 6.1: Specifications at 12V for of DC-motors used in the experiment
this comes down to 3200 counts per revolution of the output shaft. This is an unnecessarily high
resolution for the purpose of this project.
6.1.3 Electronics
To power and control the motors, drivers are required. The Pololu Dual MC33926 Motor Driver
Shield [20] was used for this purpose. It combines two drivers on one shield that can be placed on
top of the Arduino. This way the electronics are easy to install and all features are compatible with
the Arduino environment.
The drivers are rated for delivering 3A of continuous current and feedback information about the
current drawn by the motors. It also provides built-in protection for over-current and excessive
temperature.
Figure 6.1: Pololu Dual MC33926 Motor Driver Shield. Note that in the figure the shield is connected to a
Arduino Uno board
The shield is connected with a 12V voltage source for feeding the motors and additionally it can also
deliver power to the Arduino board so keeping it from drawing current from the attached computer.
The encoders are connected to the Arduino directly as well, some digital output ports were set high
or low for the purpose of providing a reference voltage to the encoders. This way, a breadboard
was not needed for the experiments.
6.1.4 Mechanical setup
All the parts, (i.e. the Arduino with driver and motors with spools and rope) are then attached to a
piece of wood. The wood adds extra weight simulating the weight off the batteries and brush. In
total the physical model weighs 1kg.

The spools attached to the motors are 3D printed and have a radius of 4cm. Also a 3D prined wire
guide was added to the top of the robot so the wire gets rolled up properly onto the motors.
The ropes were then attached to the corners of a closet, giving it a simulated roof width of 1.1m
and a length of 2 meters. The electronics were connected to a voltage source providing the motor
driver with 12V. The Arduino was hooked up to the computer using a long USB cable.
As is to be seen in figure 6.2, the motors are not connected at the same height. This is because
once the motors are screwed into place the spools couldnt get slid onto the shafts anymore. This
difference introduces a slight imbalance making the robot tilt slightly.
Figure 6.2: Test set-up of the the small scale physical model
6.2 Comparing the models step responses
6.2.1 The experiment
A good way to check the models correctness is by comparing there responses to a step function.
Because both the virtual as the physical models are controlled the same, i.e. having the same
control, they should both react the same way to a given step function.
Each of the following experiments were conducted following the same procedure. In Simulink,
a signal generator was used to create a function for both the x and y coordinates. That signal
generator then was used as setpoint value for the controller.
Because the robot needs some time to go to its begin position and stabilize, the step function was
only applied after a settling time of 15 seconds.
Afterwards, the results to the same step function can be plotted into one figure thus providing an
easy way to compare their responses.
For the PID controller only the gain value was used. Without this gain, the controller would output
voltages that are too low to turn the motor at all.
Different step functions were applied to the system.
a vertical step function in the middle of the roof
a vertical step function at the side of the roof
a diagonal step function (both x and y)
In all tests, the step functions were chosen so that the ropes were rolled up and never had slack.
This because for negative rotations the models are not equally correct. (see 5.3.2)
The conversion from polar coordinates to Cartesian coordinates needed for the the rope lengths,
leads to the fact that when a step function is applied to the controller (Cartesian coordinates) the
machine will not move in a straight way but in a rather arced and unpredictable path. This means
that if the physical model follows the same trajectory as the virtual under the same conditions, the
model is consequently accurate.
6.2.2 Results and conclusions
Vertical step function in the middle of the roof
The first experiment consists of pulling the robot up in a straight line in the middle of the surface.
This means that both motors should theoretically get the same input voltage and go up in a straight
way. However due too the nature of the controller, this is not the case and the robot moves a small
amount sideways. (see figure 6.3) This happens to both the virtual as the physical model.
More importantly, the speed at which the robot moves along the Y-axis is, in essence the slope, are
very similar. The speed of the computer model was 0,37 m/s and the real model moved at 0,41 m/s
meaning a fault of 10%
Figure 6.3: Step function applied to the Y input of the controller in the middle of the surface
Also has to be remarked that on the y axis there is a difference in static error between both values
after the settling time. The static error for the physical model is less than that one of its virtual
representation. The phenomenon was seen in every experiment. It can be explained by the fact
that the motors are a bit stronger than stated in the data-sheets.
Should this steady state error be too big an integral value can be added to the PID-controller in
the controller. However, because the error on the physical model is acceptable, it isnt immediately
necessary.
Note again, just like in the mathematical models, the origin of the coordinate system is positioned
in the top left corner with the y axis down. This is done so the coordinates on the plane are always
positive.
Vertical step function at the side of the roof
For the second experiment again the robot had to ride a straight line up, but this time on the left part
of the surface. This means that both motors get different commands thus making the path arced
and not straight.
Vertical speeds are virtually the same as for the experiment executed in the middle, giving it an
error of around 10% as well.
Figure 6.4: Step function applied to the Y input of the controller on the left side of the surface
Diagonal step function (both x and y)
The third experiment consisted out of pulling up the robot in a diagonal line in such a way that both
motors had to pull and never gave slack. This way, it could be tested if the controller still works well
when the robot is moved on both axis at the same time. Again the physical and virtual models had
about the same settling time.
Figure 6.5: Step function applied to both X and Y input of the controller
Another interesting thing is to take a look at the currents. The motor controller shield attached
to the Arduino microcontroller has the ability to measure the current flowing through both motors.
However, this feature is rather meant for measuring current when using reasonably constant loads
rather than fairly values. There the robot wobbles during its movement, the load is not constant
enough to be measured accurately.
However, some conclusions can be drawn from comparing the acquired data (see figure 6.6). Es-
pecially looking at the moment the robot arrives at its end position and stays there, the difference
in current is clearly visible. This is due to the backdrivability of the motor. Due to the gearbox ratio,
the effect of the friction in the motor and the gears get amplified and the motor only has to provide
a small amount of extra torque to keep the machine hanging still. The rest of the torque is provided
by the friction.
Figure 6.6: Current drawn by the right motor in the diagonal movement experiment.
At other positions more towards the middle and bottom of the surface, gravity doesnt even have
enough power to move the robot without the robot providing extra torque. Although this was not
planned, this a good thing because that means the robot doesnt consume energy while hanging
still. This in contrarily to when it tries to hang still somewhere at the top where it consumes a lot of
energy just for providing the torque to hang still.
6.3 Checking the limits of the system
6.3.1 The experiment
Knowing the system responds the same way the model does, there is a second way to validate the
model, namely how the system reacts on the high forces at the top of the roof. At this point the
angles of the ropes are high and the torque the motors have to deliver is the maximum torque they
can provide.
The way the models were compared was by doing a simple experiment. With a signal builder a
path was created in which the robot gets pulled straight up in the middle of the roof.
At some point higher up the surface the forces becomes too high and the robot stalls. The distance
between the top of the surface and the wire guider of the robot is then measured and compared.
This test can only be done in the middle of the surface. If the test is done at the side of the surface
the robot will start moving sideways as it tries to compensate for the lacking torque of one of the
motors.
6.3.2 Results and conclusions
The experiment was repeated three times. At each experiment the limiter (saturation block) in the
controller was set to an other maximum voltage. This way the maximum amount of torque to the
motors was reduced. The effect of this reduction is that the robot will stall at a different place making
it possible to compare the performance of the motors.
Voltage limit Computer Model Physical Model

12V 12cm 9cm
4V 30cm 22cm
2.5V 75cm 29cm
Table 6.2: Stall distance when pulling robot straight up
Its clear to see that the models act pretty similar at full power. However when the the voltage gets
reduced and the motors dont have the the full amount of torque, the differences become more
noticeable. This again leads to believe that the motors are stronger at lower voltages then their
virtual counterparts.
This also explains the smaller steady state errors encountered in the physical model. Because the
real motors deliver more torque at smaller voltage, the small change in voltage delivered by the
controller provides a bigger difference in torque and consequently moves the robot closer to its
setpoint.
6.4 Testing the braking qualities of the motors?
A last test consisted in checking how much weight the motors could actually carry without getting
any current from the source. In other words how big an influence has the backdriveability on the
robot in rest. The test showed that each motor can carry around 550 grams of weight. That is
explains why it can stall in the middle of the roof and not when its off center. This force is not
embedded in the current model but could possible be added if a more accurate model needs to be
made.
6.5 Conclusion
Throughout the test the virtual model turned out to be a very reliable representation for the robot.
The physical robot often was slightly stronger leading to better performances but for the sake of
simulating the agility and controllability of the robot this is not a problem. It is even rather positive
that the physical model has more power than its virtual representation there it would mean it could
work as a small safety factor for the results.
Another difference between the model and the proof of concept were that the motors internal
friction causes the them to brake when torque is applied at the output shaft. This means the robot
needs less energy to stay in a given position, causing both a longer battery life as less risk of
burning out the motors when stalling them.
The models could however be made better by adding this backdriveability feature and make the
motors slightly stronger but the it can be used as is right now.
Chapter 7
Implementation and Applicability
This chapter will apply the drive principle to a complete concept listing all the necessary compo-
nents. Also it will state the weaknesses of the system. Lastly is checked if the set goals are met.
7.1 Limits of the system
After the model is validated, it is possible to use the computer models to calculate the limits of the
drive mechanism. How much higher do the ropes have to be attached for the robot to reach the top
of the highest panel.
Also, it has to be able to have enough speed at each point in order to clean the panels properly.
As visible in table 7.1, the stall distance is always around 5% of the distance between the 2 attach
points. To be sure the robot can still move at a decent speed it is recommended to keep the distance
between the edges of the panels and the corners of the system at least 10% of the width.
Roof width 30 35 45
2m 5cm 13cm 14cm
3m 15cm 16cm 18cm
4m 17cm 19cm 24cm
5m 21cm 24cm 30cm
6m 25cm 29cm 36cm
Table 7.1: Stall distance for different roof settings
In addition, the system also has lower limits. The robust controller used for the experiments (see
5.3.3) is only be stable until roughly two times the width of the surface that has to be cleaned. Lower
than that, the lines are so long and the respons so different that another controller is needed in order
to move the robot accurately. However, in practice this kind of setup will not be very common.
7.2 Power consumption and battery
The model allows to get a rough estimation of the amount of power consumed by the motors. The
power consumption of the system determines the needed capacity of the battery.
44
To get these data, a relatively steep roof of 45 degrees was assumed and then a roof of 6 by 4
meters was virtually cleaned.
Afterwards the data collected by the virtual voltage and current sensors were then processed into an
average power consumption and battery capacity. This resulted in the following power consumption:
The average current found was 0.96A
The average voltage was 4.56V
This leads to an average of 4.4W
The chosen speed for the robot was 0.15m/s. With a cleaning width of around 15cm it means
it can potentially clean 81m2 per hour.
This leads to an average power consumption 370mAh (at 12V) or 4.6mAh per m2 .
There the model is not very accurate regarding current this number has to be taken as rough esti-
mate. More so because some power has to be added for the microcontroller and friction. However,
a 500-1000mAh 12V battery should have enough energy to safely provide the energy for an aver-
age sized plant. To validate this numbers, extra experiments tests should be done to measure the
power consumption more accurately.
7.3 Path
The pathcontroller sets out a route to follow by the robot. Choosing an optimal path could also
benefit the cleaning efficiency. In the literature study is already mentioned that dirt builds up espe-
cially at the lower edges. Giving them more attention as in figure 7.2 when cleaning the panels is
therefore important.
Table 7.2: Example of a possible cleaning path focusing on the lower edges.
An optimal pathcontroller should have to know the size and placement of each panel. This way, it
knows every horizontal edge were dirt can build up. However, this is in contrast with the goal of the
project to make the device as flexible and easy to adapt as possible. A way to overcome this could
be by adding an infrared distance sensor at the bottom side of the device. This way the edges can
be detected, or rather the gaps between them, and mapped in order to pass these points more.
7.4 Brush
As mentioned earlier in the literature study, there are different techniques of washing the panels.
The most important point is to keep them from getting dirty in the first place, because it gets harder
to clean them over time. The models showed that the robot is capable of moving in straight lines
thus making it possible to clean the panels while being dry. This means brushing the dirt downwards
and off the panels the way the Ecoppia E4 does it. (see 2.4.2)
However this would require a brush that rotates while the robot moves. This therefore requires an
extra motor, gears and bearings thus making the entire machine heavier. Furthermore, there is the
problem that the robot can not go over the edges without falling off or getting stuck. If a rotating
brush was to be added at the front of the robot, it wouldnt be able to clean a part the size of its own
size when approaching the lower edge with his back end. In addition, the robot would need to be
equipped with wheels in order to keep the robot from pushing too hard on the brushes to keep the
robot from having too much weight on the brushes, preventing them from turning. The wheels, in
their turn could get stuck in the gaps between the panels and could potentially damage the panels if
they are not made from a protective, and consequently, more expensive, material. Although it might
be possible, it is not worth inquiring there the goal is to make an inexpensive, profitable robot.
When cleaned wet, the brush doesnt necessarily moves towards one direction. Henceforth, the
brush doesnt have to rotate and the force applied to the brush can be higher. The robot then,
instead of using wheels, uses the cleaning surface to divide its weight on.
The brushes made to manually clean the panels (see 2.4.5) would make a good candidate. They
are often made from the same material as boat cleaning brushes and are made to both deliver
water to the bristles and more importantly, prevent them from scratching the cleaned surface.
Boat cleaning brushes are easy to find and are virtually inexpensive compared to using wheels and
rotating brushes. They are often made from Tampico. [21] It is a stiff fiber derived from the Agave
plant which grows mostly in Mexico. A drawback though is that the panels need to be wet, thus
requiring a rain sensor on the robot.
Friction is also a big factor in choosing the right brush. Gravity has to overcome the static friction
in order to start moving the robot. Finding the critical friction angle can be done with the following
formula:
tan = s (7.1)
In the this formula, stands for the critical angle and s stands for the friction coefficient between
both surfaces. Quick experiments with a similar brush on a glass mirror lead to an estimation of a
25 degree angle for the critical angle resulting in a static friction coefficient of around 0.5.
Because the robot is meant for steeper angles (30 35 ) this might suffice however finding a brush
with either lower friction coefficients would be beneficial because it could then be applied to even
lower smaller angles.
7.5 Additional electronics
7.5.1 Power
The robot has to be provided with energy. A way to do this would be by using a docking station in
one of the upper corners of the roof providing the robot with energy. The docking station would be
necessary anyway for keeping the robot of the panels when it is not shut down.
The electricity source could either be a small solar panel on the docking station or could come from
the mains. This second solution would make more sense because a solar panel would only add to
the overall price of the robot.
7.5.2 Calibration
For operation on a real roof there is a need for more electronics. In the first place, there is a need
for an onboard accelerometer. This to prevent the robot from getting stuck. A fairly inexpensive
accelerometer could give feedback about the movement of the robot. The rotary encoders cant
detect when the robot gets stuck on the way down and without this extra help, they would keep
giving slack to the ropes, risking them getting tangled or stuck on the a panel.
The accelerometer could then also give feedback on the direction and speed the robot is travelling.
This way the robot could, based on test movements, check the response. Based on this feedback,
the system can calculate how wide the plant is or even easier, explore the panel array and determine
the size on its own.
Here is an example of how this could work:
The robots docking station is positioned in the upper left corner above the panels. It is
positioned in such a way that the side of the robot is aligned with the side of the array, this
way the robot knows its as far to the side as it can get.
Then the robot, hanging with all its force on the left motor, can lower itself on just the left mo-
tor. The right motor has to give slack without knowing how much, however the accelerometer
can help with determining if the right motor is giving too little or too much slack.
When the robot reaches the underside of the panels an infrared distance sensor notices the
end of the panels and feeds back this information to the microcontroller. Now the robot knows
how long the array is.
Next the robot tries to slowly make its way across the bottom edge of the panels by combing
its accelerometer data with the feedback of the infrared sensor keeping it from falling of. At
some point the robot is exactly under the right motor and therefore at the right down corner,
it know knows how wide the array of panels is.
Calibration is now complete. These steps can be repeated if needed to more precisely
calibrate it.
This is an easy approach to calibrating the robot. Sadly it wouldnt suit to ever kind of arrangement
because it would only be applicable to squared set ups. If for example there is no panel at the top
right of the array, the robot would simply fall off. More infrared distance sensors would have to be
added in order to get through more complex situations.
A better way to calibrate would be by adding a USB port to the robot and letting the user add the
corners of the array by using a simple computer program before deploying it on the roof. This
way the encoders of the motors can be used in cooperation with an accelerometer for detecting
problems, but edge sensors wouldnt be necessary.
7.5.3 Weather
In order for the system to work properly, rain has to be detected in order to deploy the system. For
this reason, a rain sensor needs to be added to the system.
Furthermore has to be investigated wetter further weather detection sensors are needed in order
to keep the robot from deploying in bad weather. Wind, temperature and humidity sensors would
be able to detect the conditions outside. As an extra small heating elements could be strategically
placed in order to defrost some parts if needed.
7.6 Complete robot and price
Now all the parts of the robot are know, it is possible to calculate the total estimated price and
weight for the robot. Table 7.3 shows all necessary parts with their retail price and weight. For
some parts it is difficult to predict the price and or weight and are consequently estimated.
Table 7.3: Estimation of the total price and weight of the robot
All prices are consumer retail prices and include VAT. This means the robot can be produced for
a lower amount of money if it is made on a larger scale. Also, the electronics like the Arduino
microcontroller and the drivers can be incorporated into one custom made printed circuit board
(PCB). The Arduino is in fact only a development board. The actual microprocessor that drives it
only costs a fraction of the price. In the final product, only the microcontroller itself and some small
inexpensive components have to be added. The same way the accelerometer chip could also be
implemented together with the power electronics to PCB.
The other more expensive parts like motors, brushes,... will in reality be cheaper than the retail
price when bought on a larger scale. However the fact that a prototype could be developed for this
price means that producing a product and offering it for 199 Euro or lower wouldnt be unreachable
while keeping a healthy profit margin.
7.7 Calculation of return on investment
As stated earlier, the return on investment is dependent on the scale of the total panel surface,
estimated losses and energy price. The goal is to earn back the money in around two years
because this is the duration someone has to give warranty on an electronic product in the most
European countries.
For the calculation of the return on investment following assumptions were made:
The robot cleans the panels perfectly and 0% energy is lost.
The by the robot consumed energy is negligible.
There are no installation costs.
When the size of the panels is 30m2 , the size that covers an average families consumption (see
199Euro
2.3) the robot earns itself back when the plant has losses of 630Euro2 = 15, 8% per year. This is
rather a lot but isnt impossible either. This matches success level two from 3.1 and consequently
has a potential to be applicable. For every year it works longer from then on, it would return 100
euro per year.
However, for plants that are bigger then that, the investment will be returned even at lower losses.
It must be said though that the retail price of the robot would have to be even lower then 199 euro
to make it certainly worth its money.
7.8 Secondary and tertiary goals
Lastly the secondary and tertiary goals set earlier on in the project are checked. Most of the
secondary goals (see table 7.4) are met. Only the weatherproofness of the robot is questionable
there it could not be tested and would take a completed prototype to do so. It is difficult to predict
how the parts would react on cold weather but this would certainly have its toll on some of the
components like the battery. Also, wind could possibly blow the robot sideways of of the panels
disabling the robot and possibly damaging the panels and/or the robot.
Ability to clear gaps between panels

Should work on ramped roofs higher than 30
Shouldnt damage the panels
Has to be weather resistant questionable and has to be tested
Should be as small as possible
Has to be as autonomous as possible
Has to be flexible
Table 7.4: Checklist list of the met secondary goals
The tertiary, less important goals are not met. The system would probably rather not work in snowy
conditions and the system can also not clean flat surface without adding a third rope.
Changing the brushes could theoretically be possible but be rather unnecessary because it would
only work in rather wetter climates.
Ability to clean flat panels X

Ability to move snow Rather not deployable in snow
Docking station
Interchangeable brush so it can be
used in various climates
Table 7.5: Checklist list of the met tertiary goals
7.9 Conclusion
The limits of the system dont seem to be too much of a limiting factor for the feasibility of the
system. The device needs to be hooked up to a point 10% higher than the width between the side
edges of the panels.
Furthermore is the powerconsumption of the system in an acceptable range. For cleaning an
average sized plant, a battery of 500-1000mAh should suffice. The robot would also be able to
clean around 81m2 of panels meaning which is fast enough for its purpose.
The robot can potentially go fast enough to use a passive brush system. The underside of the robot
can be equipped with boat brushes keeping the panels clean without scratching them. A drawback
of this system is that the robot should be deployed when rain fell on the roof. However the cleaning
capabilities of the brush in combination with the drive system are rather speculative. Here should
be done more research regarding how well it can clean a glass surface using only the movement
of the robot.
Next a list of all the necessary parts was made in order estimate the price of the robot. A rather
rough estimation was made in order to compare the robots price with the set goals. The result
here is satisfying. Paying itself back in a timespan of two years for an average sized plant would be
possible with losses starting from 15%. Larger sized plants wouldnt need that much soiling losses
to make a profit in the third year of operation.
Also most of the secondary and tertiary goals were. One point is concerning however. It is ques-
tionable if the robot would be able to be deployed in winter conditions. Negative temperatures
could affect both motors, ropes as, not to forget the batteries. However, it is not possible to draw
conclusions on this subject without doing further experiments.
Chapter 8
Conclusion and recommendations
Both the model as the proof of concept have proved that this drive principle for moving a robot over
a roof by using tethers is possible. Unfortunately there are limitations to the system as well. First
of all the, the forces at the top of the roof make it necessary for the ropes to be attached higher
then the highest point that has to be cleaned. The computer model has proved that the attachment
points should be around 10% higher as the width of the solar plant.
Secondly there is the issue of controlling the robot. Because the response of the robot differs at
each point of the roof, the controller has to be either very robust or rather complex and adaptive
to the situation. In this paper, the robust way was tested and turned out to suffice for wide plants.
However if the robot has to go relatively low compared to its width, it starts to have more problems
with making smaller movements. Therefore, the controller should be investigated further in order to
see if a complex controller could change this without loosing the flexibility a robust controller has.
Another limitation in terms of flexibility is the feedback system. For the experiments, quadrature
encoders mounted on the motors were used. They gave a very accurate position reading but sadly
need calibration in order to know its begin position and secondly, it needs to know the width between
the attachment points in order to calculate the current position. This can partly be solved by letting
the robot park itself in a docking in one of the corners of the roof where it then loads its battery.
Further research should be done when a complete prototype is developed and it can be applied on
full-scale. The first unresearched element is the weather compatibility of the robot. Cold weather
may block the proper functioning of the robot, heavy wind may blow it of the side and snow might
block the path. Furthermore, what are the long term effects of being exposed to the elements on
the battery, motors and other components. And most importantly, are the cleaning capabilities of
the system in combination with the brush well enough to keep the panels so clean that robot earns
itself back in an acceptable time.
Although the concept is not ready to be made into a consumer product, it does show a potential
of being able to clean small surfaces in a relatively inexpensive way. Moreover, the system doesnt
require a lot of parts and is relatively easy to install. As the return on investment implies, even with
perfect cleaning capabilities and a very cheap setup the plant would need to have losses of around
15% per year in order to be profitable. This is rather high and wouldnt cover for the potential extra
cost of installing it.
Should the robot be developed further, the models can be used as a tool for simulating different
possibilities and configurations. The experiments done so far showed that they led to relatively
reliable simulations and can even be made more reliable when the backdriveability of the DC-
51
8 Conclusion and recommendations 52
motors would be added to it.

However, the system worked out in this paper would have a lot of potential flaws in terms of reliability
and this could undermine the soul of the robot. In the first place, its purpose is to keep people from
having to climb up on there roofs and risk their lives while trying to get more energy for their money.
But the moment the robot gets into trouble even once and someone has to aid it, this whole ideal
would be lost.
Therefore the conclusion is, unless every possible flaw can get ruled out, unless the price can even
be decreased and unless its cleaning capabilities turn out to be near perfect, the robot would not
be worth investing in.
Bibliography
[1] EUROBSERVER. Photovoltaic barometer, April 2014.
[2] A.R. Jha. Solar Cell Technology and Applications. Auerbach, 2010.
[3] Maxim Tabachnyk, Bruno Ehrler, Simon Gelinas, Marcus L. Bohm, Brian J. Walker, Kevin P.
Musselman, Neil C. Greenham, Richard H. Friend, and Akshay Rao. Resonant energy transfer
of triplet excitons from pentacene to pbse nanocrystals. Nature Materials, 2014.
[4] Wei-Jen Lee Jie Shi, Fellow, IEEE, Yongqian Liu, Yongping Yang, , and Peng Wang. Forecast-
ing power output of photovoltaic systems based on weather classification and support vector
machines. EEE transactions on industry applications, 48, 2012.
[5] Emmanuel Kymakis, Sofoklis Kalykakis, and Thales M. Papazoglou. Performance analysis of a
grid connected photovoltaic park on the island of crete. Energy Conversion and Management,
50, 2009.
[6] Felipe A. Mejia and Jan Kleissl. Soiling losses for solar photovoltaic systems in california.
Solar Energy, 95:357363, 2013.
[7] Miguel Garcia, Luis Marroyo, Eduardo Lorenzo, and Miguel Perez3. Soiling and other optical
losses in solar-tracking pv plants in navarra. Prog. Photovolt: Res. Appl., 19:211217, 2011.
[8] Reinhart Appels, Buvaneshwari Lefevre, Bert Herteleer, Hans Goverde, Alexander Beerten,
Robin Paesen, Klaas De Medts, Johan Driesen, and Jef Poortmans. Effect of soiling on
photovoltaic modules. Solar Energy, 96, 2013.
[9] Sanaz Ghazi and Kenneth Ip. The effect of weather conditions on the efficiency of pv panels
in the southeast of uk. Renewable Energy, 69, 2014.
[10] Paula Sanchez-Friera, Michel Piliougine, Javier Pelaez, Jesus Carretero, and Mariano Sidrach
de Cardona. Analysis of degradation mechanisms of crystalline silicon pv modules after 12
years of operation in southern europe. Progress in Photovoltaics: research and applications,
2011.
[11] Gerd Becker, Bruno Schiebelsberger, Walter Weber, Christian Vodermayer, Mike Zehner, and
Gerald Kummerle. An approach to the impact of snow on the yield of grid connected pv
systems. Proc. European PVSEC, 2006.
[12] Vreg. Het gemiddelde elektriciteits- en aardgasverbruik.

http://www.vreg.be/
info-over-het-gemiddelde-elektriciteits-en-aardgasverbruik, 2014.
53
BIBLIOGRAPHY 54
[13] Vlaamse overheid. Zonnepanelen voor elektriciteit. http://www.energiesparen.be/groene-

energie-en-wkk/zonnepanelen, 2015.
[14] Ridha Azaiz. Solarbrush - cleaner solar energy. http://www.solarbrush.co, 2014.
[15] Ecoppia. Field-proven, automatic, water-free solar panel cleaning system. http://www.
ecoppia.com/ecoppia-e4, 2014.
[16] Washpanel. Product ranges. http://www.washpanel.com/en/prodotti.php, 2014.
[17] Heliotex. Automatic solar panel systems. http://www.solarpanelcleaningsystems.

com/residential.html, 2015.
[18] R Merzouki, A.K. Samantaray, P.M. Pathak, and B. Ould Bouamama. Intelligent Mechanical
Design. Springer, 2013.
[19] Arduino. Arduino mega 2560. http://www.arduino.cc/en/pmwiki.php?n=Main/ArduinoBoardMega2560,

2015.
[20] Pololu Corporation. Pololu dual mc33926 motor driver shield users guide.
https://www.pololu.com/docs/0J55, 2015.
[21] MFC ltd. What is tampico. http://www.mfc-usa.com/tampico.html, 2003.
[22] Pololu Corporation. Catalog: 50:1 gearmotor 37dx54l.

https://www.pololu.com/product/1444/specs, 2015.
[23] MathWorks. Dc motor model. http://nl.mathworks.com/help/physmod/elec/ug/example

modeling-a-dc-motor.html, 2015.
[24] Soteris A. Kalogirou. Solar Energy Engineering: Processes and Systems. Academic Press,
2009.
[25] Martin A. Green. Silicon photovoltaic modules: A brief history of the first 50 years. Prog.
Photovolt: Res. Appl., 13:447455, 2005.
FACULTY OF ENGINEERING TECHNOLOGY
TECHNOLOGY CAMPUS DE NAYER
Jan De Nayerlaan 5
2860 SINT-KATELIJNE-WAVER, Belgium
tel. + 32 15 31 69 44
fet.denayer@kuleuven.be
www.fet.kuleuven.be

Kegel Eers

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Kegel Eers

Загружено:

Авторское право:

Доступные форматы

FACULTY OF ENGINEERING TECHNOLOGY

TECHNOLOGY CAMPUS DE NAYER

The development of a clean-

Academic Year 2014 - 2015

Door de groeiende interesse in hernieuwbare energie is de vraag naar fotovoltasche zonnepanelen

4 Drive principle selection 18

5 Modelling the robot 25

6 Validating the models 36

7 Implementation and Applicability 44

8 Conclusion and recommendations 51

2.1 PV system power output under different weather conditions. . . . . . . . . . . . . . 4

3.1 Flowchart of the methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.1 Success levels for the project with explanation. . . . . . . . . . . . . . . . . . . . . 16

4.1 Ranking table for the design selection process . . . . . . . . . . . . . . . . . . . . 21

6.1 Specifications at 12V for of DC-motors used in the experiment . . . . . . . . . . . . 37

7.1 Stall distance for different roof settings . . . . . . . . . . . . . . . . . . . . . . . . . 44

2.1 Photovoltaic energy

2.1.1 Working principle

2.1.2 Weather influences

Figure 2.1: PV system power output under different weather conditions.

2.2 Losses in PV systems

internal network (net )

system availability and grid connection network ( ppc )

Therefore, the performance ratio (PR) can be expressed as following:

Figure 2.3: Transmittance decrease due to dry residue after evaporation

2.3 Residential use of solar panels

An average Flemisch family consumes around 3500 kWh each year.

2.4 Existing cleaning robots

Figure 2.6: The solarbrush robot without its brush.

Table 2.2: Pros and cons of Solarbrush system

Figure 2.7: Ecoppia E4

Table 2.3: Pros and cons of Eccopia E4

2.4.3 Washpanel systems

Table 2.4: Pros and cons of Washpanel system

Figure 2.8: An example of a Wash Panel system.

Table 2.5: Pros and cons of a sprinkler system

2.4.5 Manual cleaning

Table 2.6: Pros and cons of manual cleaning

Figure 2.9: A person cleaning PV panels manually with waterfed pole.

Figure 3.1: Flowchart of the methodology

Level Size Loss Explanation

Table 3.1: Success levels for the project with explanation.

Ability to clear gaps between panels

Should work on ramped roofs

Shouldnt damage the panels

Has to be weather resistant

Should be as small as possible

Has to be as autonomous as possible

Ability to clean flat panels

Ability to move snow

Interchangeable brush so it can be used in various climates

Drive principle selection

Horizontal Rail with Vertical Brush

Horizontal Rail with Horizontal Brush

Two Parallel Tethers

Two Corner Tethers

Figure 4.3: Conceptual representation of designs using two or four tethers.

Four Corner Tethers

4.1.4 Selection Process

Cost: An intuitive guess of the relative price.

Table 4.1: Ranking table for the design selection process

4.2 Choosing the motors

Table 4.2: Pros and cons of DC and Stepper motor.

4.2.1 Characteristics of the DC motor