Előadás

Deep Learning techniques applied on meteorological data for weather nowcasting purposes

Andrei Mihai

BBTE

Beyond Artificial Intelligence Research: a New Innovation and Industrial Domain

2025. november 7. 13:00Babeș-Bolyai University, Central Building, Aula Magna · Kolozsvár · Románia

Átirat · 25:14

Yeah. All right. Okay. So, thank you for the introduction again. I'm Andre. So I

Átirat · 25:19

will present a bit some of the work that we did uh with uh in the field of

Átirat · 25:25

methology. Um so this work that we did was part of the women project uh a

Átirat · 25:35

project that is a collaboration between the our university and two uh

Átirat · 25:41

administration methological administration the one from Romania anime and the Norvigial uh methological

Átirat · 25:47

administration. [clears throat] Right. So uh what I will present to you is what uh we did a bit

Átirat · 25:54

what data we worked with what models we have created um for the purposes uh well now casting

Átirat · 26:03

I will explain a bit uh now casting purposes um

Átirat · 26:08

yeah so the goal well one of the goal of the projects and uh some that we worked

Átirat · 26:14

in is weather now casting so the idea with uh now casting is to try to predict

Átirat · 26:22

the weather for a very short term. In general, it's defined at zero to six

Átirat · 26:27

hours predictions. We work to less than one hour. So very very short term just trying to predict uh as accurately as

Átirat · 26:34

possible the weather next time. Well uh actually the purpose of our work was to

Átirat · 26:41

try to predict radar data. Okay. So how this works is that uh methologist

Átirat · 26:47

operational methologist they gather the data from the radars right and they try

Átirat · 26:52

to use that data that they have to understand what is happening and make

Átirat · 26:58

their predictions right so the purpose for the now casting models that we're trying to create is to kind of predict

Átirat · 27:07

what the radar would read in slightly in the future right so the purpose of this

Átirat · 27:12

is to help them the then a meteorologist would see oh okay the AI predicts that

Átirat · 27:18

this will be the situation in I don't know five 10 20 minutes so then they can

Átirat · 27:24

make better predictions if there will be bad weather or make warnings and so on

Átirat · 27:29

so this is the idea that we worked on okay so here I'm just trying to explain

Átirat · 27:36

a bit the data that we work with so uh yeah in AI research a lot has to to

Átirat · 27:44

actually to understanding the data and learning how to work with it uh other

Átirat · 27:50

than just creating the models themselves. So uh the radar data whether

Átirat · 27:55

uh radar actually gathers data in in a certain area on different levels right

Átirat · 28:02

on multiple levels and what we received as data are multiple matrices right

Átirat · 28:09

where uh each point is a physical location on yeah geographical location

Átirat · 28:16

and it says uh the level there of the data uh and yeah I yeah you can see the

Átirat · 28:23

example there. Uh so each matrix represents one different

Átirat · 28:29

level of altitude basically. Uh so we get all of these data every five or six

Átirat · 28:36

minutes. So actually the radar gathers a lot of data. We had to understand how to uh use that data. So uh also the radar

Átirat · 28:44

actually creates a lot of different types of data about the wind about the um

Átirat · 28:50

weather droplets and about everything else that it can find in in the

Átirat · 28:57

atmosphere. What we used is the reflectivity which is

Átirat · 29:02

well basically the water droplets how big is the water in the atmosphere

Átirat · 29:07

something like that. Um so we have this is the data that we have. We have reflectivity which is very useful for

Átirat · 29:14

materologist. We have it on different altitude levels and we have it every five or six minutes. So what do we do

Átirat · 29:22

with it? Uh so first of all right we we can see it like these kind of cubes or

Átirat · 29:28

images where instead of having red green blue we have the each altitude level and

Átirat · 29:35

we have these images for each time step time step right every five or six

Átirat · 29:41

minutes. So this is all the data that we have and what we want to do with this

Átirat · 29:46

is to try to predict the same thing that would be in the future like the same set

Átirat · 29:52

of data. So we have the data that is I don't know the radar just gathered now

Átirat · 29:58

what will the data give well what will the radar give in the next 5 minutes and

Átirat · 30:04

this is the very basic case but then we have like let's use all the data in the last 20 minutes to try to predict for

Átirat · 30:11

the next 5 10 15 minutes and so on. So um this is

Átirat · 30:18

um yeah the the final the models that we created this yeah this is what they are

Átirat · 30:25

trying to to do in in general. All right. So now I will go a bit to how we

Átirat · 30:32

created some of the models. One of them is this next now model that we uh created. It is um based on the ResNet.

Átirat · 30:41

So uh it is um well a bit more of a uh

Átirat · 30:48

it's a deep learning ResNet was a deep learning model that was one of u the best models at the time. Uh but yeah

Átirat · 30:56

it's before this foundational models and all the other new advancements. Um so

Átirat · 31:02

for example how we uh tried to create a better uh thing for our purposes. So for

Átirat · 31:10

example some how a research works as an example here in AI right we have uh this

Átirat · 31:17

network model is has multiple layers right and all these layers well in

Átirat · 31:22

resest case the they have some kind of blocks that use convolutionals in some way and what we try to do is to make

Átirat · 31:30

those residual blocks better by using uh grouping convolutions instead of simple

Átirat · 31:35

convolutions because yeah that's uh increased performance and so on. So

Átirat · 31:40

yeah, the AI research can be done also in these smaller steps. You take a model and you try to update it a bit and make

Átirat · 31:48

it better for your own case and that's what we did. Um

Átirat · 31:53

yeah so how this model worked in general you can see that it made the smaller uh

Átirat · 31:59

the image it made it smaller with uh yeah so the radar data as I said we can see it as an image with multiple

Átirat · 32:06

channels instead of red green blue it has then those layers on of altitude and

Átirat · 32:12

these models try to learn a bit different things about the image about

Átirat · 32:17

the shapes and I don't know the edges of and whatever things and we make it

Átirat · 32:24

smaller and then we make it bigger to try to recreate whatever it should be in the next 5 minutes or 20 minutes or what

Átirat · 32:31

what we trained the model on. Right? So this is the idea of how we trained it's

Átirat · 32:36

an example of a model that we made and how we created it based on a different

Átirat · 32:42

model that existed and uh yeah for example this is I don't know how well you can see the

Átirat · 32:49

the prediction but yeah it's some of the results with the comparison with

Átirat · 32:54

prediction and observation I don't know if you can see but u yeah the one of the

Átirat · 33:00

problem with the models in general that you have is that uh predict icting higher values of this reflectivity is

Átirat · 33:06

really hard. Um, okay. And this is prediction at 15 minutes. It's similar.

Átirat · 33:13

Okay. I I'll just go to the next part. So, this is another model.

Átirat · 33:18

Okay. I guess you can see really well what the model is about. This is a model that we have tried to create ourselves.

Átirat · 33:24

Just some convolutional uh layers um one after the other. The the innovation here

Átirat · 33:32

was that we didn't try to make the image smaller. We tried to make it uh to keep

Átirat · 33:37

it at the same size and that made it to learn a bit better the shapes and well

Átirat · 33:45

so on. It learn a bit better how rather uh weather and reflectivity changes,

Átirat · 33:52

right? Um but this also increased a bit the computational cost. So we had to

Átirat · 33:57

make some adjustments like using different types of layers to make it more better. All right. Uh so this uh

Átirat · 34:04

new model the septcom model is what we used to predict uh to try to predict up

Átirat · 34:11

to 45 minutes. But as I said since we

Átirat · 34:16

don't make the image smaller anymore it is much more computationally expensive. So instead we tried to think okay how to

Átirat · 34:24

actually predict all these um multiple time steps into the future. So

Átirat · 34:31

we thought what we if we train three models, one to predict five well five

Átirat · 34:37

minutes into the future, one to predict uh approximately what is it 15 minutes and one 25 minutes and then try to

Átirat · 34:44

combine the results of all these models in order to have a much more linear uh

Átirat · 34:50

to have all the the steps that we need 5 10 15 and all the other steps in

Átirat · 34:56

between. Right? So the idea was to have fewer models and combine them to have

Átirat · 35:03

better uh well to use fewer resources because we didn't have so many resources

Átirat · 35:08

to train all these models. Uh so here is yeah I've shown there how we combined

Átirat · 35:15

the models we have we had some different ideas and tried some of them and there

Átirat · 35:21

are some results on combining u using our threebased models then we

Átirat · 35:28

combine them in different ways to predict the data at u up to 45 minutes into the future.

Átirat · 35:36

Right? So the idea is that whatever some model predicted then we use that in a

Átirat · 35:42

different model to predict even further into the future and so on. Uh I won't go

Átirat · 35:47

very much into the details how we combine them because I don't know if that yeah okay um so now I show here

Átirat · 35:55

some performance evaluation metrics. Uh these aren't really important right now

Átirat · 36:01

exactly what the metrics are and so on. What I wanted to show you here is how many of them there are, right? So, uh,

Átirat · 36:09

one of the things that we actually had problems with and this can be

Átirat · 36:14

interesting in many fields of AI, I guess, is how do you actually evaluate

Átirat · 36:19

how good of a job you did, right? You you have to use some kind of mathematical formula to try to say,

Átirat · 36:26

okay, this is good or it's not as good. And that is not that easy, right? So

Átirat · 36:32

there are uh we have here three sections of formulas. One is the two

Átirat · 36:39

sections like root mean squared error or even air squared. There are some that are very used in computational

Átirat · 36:46

um science right uh but these aren't really made for weather forecasting or

Átirat · 36:53

for now casting and aren't really that good for that case. So we researched

Átirat · 36:59

some of them for that case and we saw that uh actually used in forecasting is

Átirat · 37:05

a lot of classification metrics and well we didn't do classification but we tried

Átirat · 37:10

to change how we measured and how we have validated our data to try to fit

Átirat · 37:16

some classification metrics in there to be able to compare with the literature and and to other uh elements. So like

Átirat · 37:25

there are some scores here that are more specific to methodology but yeah there are classifications which we didn't

Átirat · 37:31

really do and then we tried to figure out some image processing metrics. Okay,

Átirat · 37:36

we try to figure out there are all these kind of image processing metrics that we have if we consider our radar data as

Átirat · 37:43

images maybe those are better and we try to find out which one would be better

Átirat · 37:50

but even that proved to be quite difficult. We well we actually made a

Átirat · 37:56

small experiment trying to figure out uh which would be the best image processing

Átirat · 38:01

metric um by asking multiple methodologist what

Átirat · 38:06

their idea was showing them some images and so on and trying to find it on a rank based and so this is how we find

Átirat · 38:13

these I don't know if you see visual information fidelity metric u right so

Átirat · 38:19

the idea is that we have all these kind of metrics that all shows us something about what we did but it's really hard

Átirat · 38:26

to you know find out. So here are also the numbers again the numbers and everything

Átirat · 38:32

doesn't really matter. What it matters is what what do we do with these numbers and actually just just having these

Átirat · 38:38

numbers here doesn't really tell you did we do a good job or not

Átirat · 38:44

you know just looking at them you can't really see that and uh what you can do

Átirat · 38:49

is try to compare to related work right what did others do what do other

Átirat · 38:54

methologist human methologist actually do if some there are some measures but

Átirat · 39:00

measures by of Human predictions are a bit rarer. It's hard to actually get

Átirat · 39:07

data of how human predicted and compare the data to your uh model AI model,

Átirat · 39:14

right? Um and even other

Átirat · 39:19

I don't know other groups and other research labs that did uh this kind of

Átirat · 39:24

research on methology they have their own metrics that they used. So we try to

Átirat · 39:30

if they used a different metric than us then we couldn't really compare the results and even we could compare the

Átirat · 39:36

results it's really hard to say is it really that much better or not because

Átirat · 39:42

they use different data right so some data might be easier to predict than other they use different methodology and

Átirat · 39:48

different things. So sometimes it's really hard to to compare different kind

Átirat · 39:54

of what you did to others and uh this is a bit what I wanted to highlight here.

Átirat · 40:00

We have this data, we have these results, and these results after all the

Átirat · 40:05

analysis that we could do tells us that we did kind of a good job, right? Uh

Átirat · 40:11

compared to the literature, they're not worse, not also much better. But again, all this comparison is a bit yeah, it's

Átirat · 40:19

not quite clear because again, it's not the same methodology, not the same data. It's hard to really do a good job at

Átirat · 40:26

this. On the other hand, what these results can show you is tell you a bit how your models work. So again, I don't

Átirat · 40:33

know if you can see this good in all these numbers, but uh one thing is that

Átirat · 40:39

it really showed us that the model have a much higher problem predicting higher

Átirat · 40:45

values. So higher reflectivity values again reflectivity means how much water is in the atmosphere, right? And

Átirat · 40:53

meteorologist really care about that the most because when there's a lot of water in the atmosphere, there is where the

Átirat · 41:00

storms are and when everything else is right and they care much more about that

Átirat · 41:05

than when there are just a few clouds or there is nothing. So

Átirat · 41:10

the fact that we actually show that these models work much worse when there is higher level of reflectivity told us

Átirat · 41:18

that yeah there is something that is missing. So it it actually showed us a bit just you know something about how to

Átirat · 41:26

try to conduct the uh the experiments further right. Um okay

Átirat · 41:34

and it also showed us uh yeah that it is harder to predict at the actual weather

Átirat · 41:40

that the methodologist wants than the one that we have well than everything

Átirat · 41:45

right. All right. So, uh this was about the

Átirat · 41:50

results here. I will show you some uh visualization. I don't know if you can see them well of the observation and

Átirat · 41:58

prediction. Uh but more than that, I think it will be a much more interesting to see a video, right? So on the left

Átirat · 42:04

side here, you can see the uh actual data, the real data, and on the right

Átirat · 42:10

side, you can see our model predicting 20 minutes in advance. So you can see how well it did. Uh, right. Uh, this

Átirat · 42:18

this it should be a video. I don't know if it why it doesn't start,

Átirat · 42:24

but for for some reason it's it doesn't want to work, but uh yeah.

Átirat · 42:33

I Well, I guess that's it then. [laughter] Uh, right. Thank you.

Átirat · 42:43

[applause] Thank you. Let's let's see if the the uh

Átirat · 42:48

just a second if it were would work. I don't know, but it it doesn't, right?

Átirat · 42:54

It doesn't seem to be a video. No. No, it's a PDF. No. Okay.

Átirat · 43:02

You tried and it worked. Okay. It didn't. But it's fine. Okay. No problem. Okay. Thank you. Thank

Átirat · 43:07

you. So

Átirat · 43:12

uh so we uh arrived to the last presentation uh for today uh which is uh

Átirat · 43:18

by Leah Chhatau uh who is the head of data mining research group here in

Átirat · 43:23

University of uh uh Babes Bon Bes University as a professor and uh as I

Átirat · 43:32

said he really helped a lot at this uh conference to be a possibility and as we

Átirat · 43:40

see now a success. So

Átirat · 43:55

where's the desk?

Átirat · 44:14

Uh-huh.

Átirat · 44:23

Control L. Okay. So, the floor is yours.

Témák

MI-támogatott

Gépi tanulás Adattudomány

Összefoglaló

MI-támogatott

Andrei Mihai a radaradatokra épülő rövid távú időjárási nowcasting gépi tanulási megközelítéseit ismerteti. Bemutatja, hogyan képezik le a több magassági szinten, 5–6 perces időközönként érkező reflektivitási adatokat többcsatornás képszerű reprezentációra, majd ResNet-alapú és saját konvolúciós modellekkel 5–45 perces előrejelzéseket készítenek. Jelentős figyelmet szentel az értékelés problémájának: a meteorológiai, képfeldolgozási és klasszifikációs metrikák eltérő képet adnak, miközben a modellek éppen a nagy reflektivitású, viharos helyzeteket jósolják nehezebben.

Videó navigáció

Fejezetek

MI-támogatott