Thank you for the kind introduction and I want to start right off with a PRS slide and

the work that I've done in basically the last year.

And there's one published or some published publications and there's one publication

under preparation and this is the one that I will be talking about today.

So this is kind of work in progress and you will notice at some point that it's like that.

And I hope to get your feedback on that and incorporate it into the further work.

So let's start with the motivation.

Why do we talk about PV power estimation?

And the main question that we want to solve basically is we want to judge whether a solar

module is still okay, whether we still would want to use it in the field because it's performing

well or if it's not like that.

And let's take a step back and I want to briefly outline what you can see here because probably

not all of you are familiar with this kind of modality.

We're using an electro-luminescence image here which means that we apply a current to

the module and then you get this kind of emitted light in the near infrared spectrum only for

the regions of the module that are active.

So the regions that are basically connected to the connections of the module.

And all regions that are disconnected appear black here.

So we roughly assume that these regions cannot contribute to the power production of the

module.

So what we want to do, we want to work toward an automated analysis of PV power plants and

we want to work towards contactless assessment of module power which is not, we're not yet

that year because that is not contactless.

So we still have to contact the module to apply the current but we're working on photoluminescence

imaging as well where we only excite the module by applying light to it and then we can have

a total contactless assessment of modules.

So let's briefly talk about the physics that is behind all that we see here and the main

thing that we discuss about today is the voltage current curve and what we want to estimate

is this point which is the maximum product of voltage and current which gives us the

maximum power, so called maximum power point the module can produce and in an ideal case

we roughly can assume that this is proportional to the active area, to the area that is not

dark in the module image.

But reality is different and this is what comes in here because there might be parts

of the module that are shaded, parts of the module that are defect and in these cases

it's not only these cells that do not contribute to the power production because the cells

are connected in series and if one of these cells doesn't work anymore the complete series

of cells doesn't work anymore and this is because there is a bypass diode here which

if one cell doesn't work anymore bypasses the current through that subdivision of the

module because otherwise this subdivision would, especially the cell that doesn't work

anymore would heat up and maybe damage the module.

So this assumption that it's proportional to the active area doesn't hold in practice

and therefore we cannot just use a very simple approach and we try to apply deep learning

to the topic.

So roughly introduce the approach that we have.

We have a data set of 719 electron luminescence images with corresponding IV curves.

So this is a very very small data set and we will see in the results that we are not

into the limit of data so it's working but I assume that it will work better if we get

more data.

We predict the power relative to the nominal power of the module because we have different

modules there with different nominal powers so we don't predict the absolute power but