WEBVTT

00:00.000 --> 00:13.840
Hi everyone, thanks for coming. This is my second year at Foss Dam, this is my first time

00:13.840 --> 00:19.240
talking, so thank you very much for listening to me. How's everyone feeling Friday

00:19.240 --> 00:29.440
afternoon, can I have a Sunday afternoon, sorry, no no, yeah, okay, I'll try and

00:29.440 --> 00:35.600
I'll try and keep the energy up. I'm Peter Dodford, I'm head of technology at Open Climate

00:35.600 --> 00:40.400
Fix and it's quite a long title, but I'm going to talk to you a little bit about what

00:40.400 --> 00:48.040
we're up to and some of the challenges we have and I'm going to try and do a little

00:48.040 --> 00:55.880
live demo for like today's a good excuse to do a live demo. Yeah, so as I said, I'll give

00:55.880 --> 01:02.320
a little introduction into what Open Climate Fix are and what we do, why we've focused on

01:02.320 --> 01:09.800
solar forecasting, we do some other things as well, but mostly solar forecasting and are kind

01:09.800 --> 01:15.520
of flagship thing, cook quarts solar a little bit more and kind of we'll get to the end

01:15.520 --> 01:23.440
on some new things that we want to do. I'm going to try and talk for 10 minutes and then

01:23.440 --> 01:30.400
there'll be some time for questions as well. So Open Climate Fix, we have founded five

01:30.400 --> 01:38.120
years ago, I haven't seen this gift thing before, so I might get distracted and there's

01:38.120 --> 01:42.360
a few people you'll see in the pictures around the room here today. We're not for

01:42.360 --> 01:50.280
profit, we're focusing, we're doing open-source solutions, we're focusing on AI solutions

01:50.280 --> 01:58.360
to help decarbonise the grid. Various experience across the board are two founders pop

01:58.360 --> 02:06.480
up there and I think some of you might know them. It's 15 of us in London, we've got one

02:06.480 --> 02:12.400
person in India as well and that's some projects we've got out there. I'm actually a

02:12.400 --> 02:21.040
Manchesterer and I end up taking the train quite a lot down to London. So why solar,

02:21.040 --> 02:25.800
someone today said the world is on fire and this sign makes me think, can I borrow that

02:25.800 --> 02:33.200
sign? I think it's 10 minutes left. I think if that's the doomsday clock, we've got to

02:33.280 --> 02:44.960
do stuff, we've got to make stuff happen. We focus on solar. I think this maps quite powerful,

02:44.960 --> 02:49.600
it kind of shows what's the cheapest source of energy in the world and once you get to

02:49.600 --> 02:56.000
2030, it's basically solar everywhere, apart from maybe like Norway or Sweden where I'm

02:56.000 --> 03:01.520
pretty sure that must, oh, I thought it might be hydro, but yes, so there's getting cheaper

03:01.600 --> 03:08.800
and there's going to be more and more of it around. You can see this already happening. There's

03:08.800 --> 03:14.800
loads of solar being built. There's loads of wind being built. My career started in the wind industry

03:14.800 --> 03:22.320
so I'm still a little part of me that I want to see the wind carry on going. Yeah, this is from

03:22.320 --> 03:32.160
member who we sometimes share an office with, so I trust them. And that's great. We're building

03:32.160 --> 03:37.920
loads of solar loads of wind, but whether suddenly it starts to become a problem is going to be

03:37.920 --> 03:41.760
lots of grids around the world that are now dependent on the weather systems that are coming

03:41.760 --> 03:47.600
around. We're getting big fluctuations in energy. It's so different from having a coal power station,

03:47.680 --> 03:54.080
having a gas power station. So we focused on, that's the reason why we focused on solar.

04:01.040 --> 04:09.120
So our main thing is, is quartz solar. This is a UK solar forecast. You can scan that and it

04:09.120 --> 04:16.320
should log you onto the app and you'll take it in if you want to. We've been taking lots of

04:16.320 --> 04:24.480
satellite images, lots of weather data, training our ML model called PVNet. And then we show it

04:24.480 --> 04:29.920
as an API. There's a nice front end that you can log into and see. We're delivering that to NISO.

04:29.920 --> 04:37.520
So that's the UK TSO. And there's a few other customers as well that we use there.

04:38.480 --> 04:44.000
We believe in this probabilistic values as well. It's never certain

04:44.480 --> 04:52.000
and how the users use that probabilistic values as important. NISO of USTAR,

04:52.640 --> 04:57.280
solar forecast in their demand forecast and showing the kind of an 8% improvement in accuracy.

04:59.600 --> 05:06.160
And they've also used it in a reserve setting project where it can be shown to save

05:06.160 --> 05:11.280
100 megawatts of reserves. So you don't have to have that 100 megawatts on. And that can

05:11.280 --> 05:16.720
you can start to like calculate rough CO2 figures with those of those amount and we think it's

05:16.720 --> 05:24.320
like 200,000 tons of CO2 year. For me, that's the so important. When I joined maybe three or four

05:24.320 --> 05:28.960
years ago, it was always like, oh, we're going to do a solar forecast. But what's what's the impact?

05:28.960 --> 05:33.920
What's going to happen? And starting to get those numbers out is important. It's impactful.

05:34.720 --> 05:36.480
We can then work on the right things.

05:41.840 --> 05:51.920
Something we've been working on recently is a Manchester prize. And that is to specifically

05:51.920 --> 05:56.720
forecast satellite images. And then we put the predicted satellite images into PV net.

05:59.680 --> 06:04.560
There's just a lot of data with all these satellite images. There are every five minutes

06:04.560 --> 06:08.160
and getting them all together and getting them in the right way and using them.

06:08.720 --> 06:15.600
It's shown we've given a five percent improvement by using that method. So if I

06:15.600 --> 06:20.240
done ML engineer called James, he's been working on that for a year and it's, you know,

06:20.240 --> 06:29.200
now it's starting to pay off, which is great. They've got a video of that.

06:31.280 --> 06:37.600
So this is the true satellite images and then the cloud forecasting ones on the right

06:37.600 --> 06:46.320
hand side. And I love looking at them. They're fascinating. And you see the different weather,

06:46.320 --> 06:51.520
patterns and different things. And probably notice it's a bit more blurry on the right hand side.

06:51.520 --> 06:56.800
And that's definitely one of our challenges is, you know, do we not want it blurry? Does that

06:56.800 --> 07:02.640
actually capture some sort of uncertainty? Yes, it's an interesting problem to have.

07:08.560 --> 07:16.480
Oh no. There we go. So here's some of the challenges. It's been data, the best definition

07:16.480 --> 07:23.040
I heard for big data was like, it's not on your laptop. So we've just got loads of it and we always,

07:23.040 --> 07:28.080
you know, every, we fill it up and we have to remove stuff and make sure we've got enough and

07:29.680 --> 07:35.040
it's stuff and soul is somewhere here is managing all of that and it's, yeah, it's a lot of work to

07:35.840 --> 07:42.080
do. A lot of that data isn't open source, the weather, the satellite, but the life satellite.

07:43.280 --> 07:50.080
And that makes it hard for the community to kind of take part and we want the community to take part.

07:50.080 --> 07:56.480
So we'll come back to that at the end. Training these models is hard to take a lot of power.

07:56.480 --> 08:01.040
There's a lot of data that we have to get in the right format and it takes time.

08:01.040 --> 08:07.360
And I mentioned about blurriness and, yeah, that's kind of one of the other challenges we've had.

08:16.320 --> 08:21.200
I'll do a quick demo. It was the baiting of I have time.

08:21.680 --> 08:31.360
So we do a site level forecast as well. So we've been doing a national one but we're also

08:31.360 --> 08:40.800
do a site level. This is completely free and open source and anyone can use it and who knows

08:40.800 --> 08:54.320
is Python here. But I can just paste that and this is where we are right now in Brussels and

08:55.360 --> 09:03.600
it should be able to just do that and it will produce a solar forecast of here right now for the

09:03.600 --> 09:09.120
next 48 hours. That's taking in weather, weather, open source, weather, kind of input.

09:12.640 --> 09:19.200
If you slightly change some inputs then you end up with a few lines and maybe that gives some

09:19.200 --> 09:24.560
uncertainty and things. I think I run that about an hour ago while sitting in the audience.

09:25.280 --> 09:32.240
We launched this last year at Boston. Right here you have part of that talk and

09:33.760 --> 09:39.520
it's great. Get him, try it out. Let us know. Let us know if it works. Should work all over the world and

09:40.880 --> 09:54.000
yeah, it's cool. So I mentioned that problem of the weather data as private. We can't share that

09:54.000 --> 10:01.920
with open source contributors. That makes it hard to develop models together. We've opened a

10:01.920 --> 10:07.920
new project called Open Data PV where we're trying to train that PV network, PV net on completely

10:07.920 --> 10:14.160
open source, weather data. So GFS and the freeze, MWF. We want to expand. We've done the UK,

10:14.160 --> 10:18.800
we do some work in India but we want to expand everywhere. We want to produce a solar forecast

10:18.880 --> 10:28.720
for lots of countries around the world. We want people to be involved and we need data engineers

10:28.720 --> 10:34.560
and ML engineers and all the way along the pipeline. So this will almost definitely be a G-SOP

10:34.560 --> 10:42.960
project so hopefully lots of other people are being involved as well. We re-wrote our ML pipeline

10:43.040 --> 10:51.120
because the old one we had was really hard and really bad. It's been great to just do it

10:51.120 --> 10:55.920
and get feedback instantly from the community. We've got more contributors to the new one than the

10:55.920 --> 11:02.960
old one already and it's only been alive for two months or so. So it's worth redoing it and we're

11:03.040 --> 11:15.360
excited by that. We've been trying to do some office hours so people can join like a live stream.

11:15.360 --> 11:23.600
We just talk and ask questions and MLE's been organizing that so thank you MLE. Just investing

11:23.600 --> 11:31.760
in the community trying to get people talking and working together. I said I don't need to talk

11:31.760 --> 11:37.440
for 10 minutes. We can still 11 minutes. So we've got some time for questions if anyone has

11:43.440 --> 11:54.320
a model architecture. Oh yeah so the question is what kind of model architecture for this one?

11:54.320 --> 12:02.000
Paving it. It's a late fusion multimodal model so there's some CNN layers and it kind of all

12:02.000 --> 12:06.720
comes together basically. We've tried to try a few other different ones and there's...

12:11.680 --> 12:16.960
Yes we've tried on some very simple. Oh let me repeat the question. Can I expand on the other ones

12:16.960 --> 12:23.440
that we've used? We've we've you know we start with XT Boost and something like that that just

12:24.320 --> 12:30.720
can't give the performance basically. We've tried some I think we've tried some transformers

12:31.280 --> 12:37.200
quite similar to what we have at the moment. There's a few of it. It's almost like tweaking

12:37.200 --> 12:42.240
different things and in some ways doing these things actually give a bigger impact than tweaking

12:42.240 --> 12:48.960
that model by you know forecasting or you know using more weather data. So it's a little bit

12:48.960 --> 12:55.520
of like where do we invest time in? Any other questions?

12:56.400 --> 13:05.280
Okay so once you know solar and power forecast how do you use that information?

13:07.600 --> 13:17.280
So you get down to the type of money. So I mean what is advantage how I guess the advantage of

13:18.160 --> 13:25.680
yeah so I guess there's from a house or a repeat question. How do people actually use the solar

13:25.680 --> 13:33.840
forecast like why why why why are we bothering doing it? There's from a good stability point of

13:33.840 --> 13:39.280
you you need a good you need to know you know roughly where the solar is going to be in the next few

13:39.280 --> 13:44.960
hours so you can plan to put on other other plants and things like that otherwise you have to have

13:45.040 --> 13:51.040
a lot of reserve, a lot of spinning reserve, a lot of plants just being at 50% just in case it changes.

13:51.040 --> 13:55.280
So if you can provide a better forecast you can have less of those reserve power plants.

13:57.120 --> 14:02.480
On a domestic level if you know how much solar there is going to be you know you plan how much

14:02.480 --> 14:07.680
I'm going to charge my battery overnight how am I going to when are we going to charge my EV thinking

14:07.680 --> 14:12.080
about dynamic prices and things like that. Does that make sense?

14:16.080 --> 14:18.080
Sorry we can talk later though.

14:27.120 --> 14:30.320
That's it the question is how many people are using it or who's using it?

14:30.640 --> 14:39.120
Yeah I mean we have a few yeah Nissan using it that's all public the fielder kind of bigger players

14:39.120 --> 14:48.480
so using it the the site level one. We've only recently it's quite hard to track right so it's more

14:48.480 --> 14:53.520
kind of like any open source projects trying to you know gauge those stories and understand

14:53.600 --> 15:02.640
what how people find it useful so yeah times up but I'll be around and you know can talk more

15:02.640 --> 15:06.640
definitely thank you everyone