WEBVTT - Here's Why AI is Changing How We Predict The Weather

0:00:02.400 --> 0:00:06.760
<v Speaker 1>Bloomberg Audio Studios, podcasts, radio news.

0:00:08.600 --> 0:00:11.080
<v Speaker 2>I'm Stephen Carroll and this is Here's Why, where we

0:00:11.119 --> 0:00:13.240
<v Speaker 2>take one news story and explain it in just a

0:00:13.280 --> 0:00:15.800
<v Speaker 2>few minutes with our experts here at Bloomberg.

0:00:21.440 --> 0:00:24.720
<v Speaker 1>Severe weather is making its way east after deadly storms

0:00:24.840 --> 0:00:25.840
<v Speaker 1>ravage the South.

0:00:26.200 --> 0:00:28.319
<v Speaker 2>It's a fact that we will see due to the

0:00:28.320 --> 0:00:33.360
<v Speaker 2>climate development, more events, more severe event The weather, as

0:00:33.400 --> 0:00:37.239
<v Speaker 2>always has a huge impact. Most shopping is still done

0:00:37.320 --> 0:00:42.159
<v Speaker 2>in person. I'm singlehandedly powering the gardening section of retail

0:00:42.280 --> 0:00:46.120
<v Speaker 2>sales data. Frankly, it's the center of so many conversations,

0:00:46.159 --> 0:00:49.599
<v Speaker 2>but also a key economic driver. A cold snap can

0:00:49.680 --> 0:00:53.000
<v Speaker 2>drive up energy demand. Sunny days can send shoppers out

0:00:53.000 --> 0:00:56.160
<v Speaker 2>to buy new clothes. Then there are extreme weather events

0:00:56.200 --> 0:01:00.120
<v Speaker 2>that can damage infrastructure, disrupt supply chains, and spark mass

0:01:00.160 --> 0:01:03.600
<v Speaker 2>of insurance claims. As weather gets more volatile in many

0:01:03.600 --> 0:01:06.360
<v Speaker 2>parts of the world, knowing what's on the way is

0:01:06.400 --> 0:01:10.559
<v Speaker 2>even more important. Technology has helped to improve the accuracy

0:01:10.600 --> 0:01:14.080
<v Speaker 2>of forecasts. So what does the explosion and artificial intelligence

0:01:14.200 --> 0:01:18.319
<v Speaker 2>mean for the science. Here's why AI is changing how

0:01:18.360 --> 0:01:23.440
<v Speaker 2>we predict the weather. Our weather reporter Mary Hoy joins

0:01:23.440 --> 0:01:25.840
<v Speaker 2>me now for more, Mary, before we get to the

0:01:25.840 --> 0:01:28.399
<v Speaker 2>effect of AI. Can you just bring us up to speed.

0:01:28.400 --> 0:01:31.160
<v Speaker 2>At how good have we gotten at being able to

0:01:31.240 --> 0:01:31.960
<v Speaker 2>predict the weather?

0:01:32.319 --> 0:01:35.959
<v Speaker 1>We've gotten very, very good. We've made incremental but tremendous

0:01:35.959 --> 0:01:39.280
<v Speaker 1>progress in the past decades since the very first computer

0:01:39.480 --> 0:01:42.440
<v Speaker 1>based weather forecast was made in nineteen fifty and that's

0:01:42.480 --> 0:01:47.560
<v Speaker 1>all thanks to various technological and scientific advancements of scientists

0:01:47.600 --> 0:01:50.720
<v Speaker 1>have made around the world. That's allowed us to divide

0:01:50.760 --> 0:01:54.560
<v Speaker 1>the world into smaller grids and then also divide the

0:01:55.000 --> 0:01:57.440
<v Speaker 1>air above us into more and more layers, and all

0:01:57.440 --> 0:01:59.240
<v Speaker 1>of that means that we can look at the weather

0:01:59.400 --> 0:02:03.400
<v Speaker 1>in more detail, which then increases our ability to see

0:02:03.400 --> 0:02:06.200
<v Speaker 1>further into the future. So you know, a five day

0:02:06.240 --> 0:02:09.120
<v Speaker 1>forecast now is as accurate as a three day forecast

0:02:09.280 --> 0:02:12.000
<v Speaker 1>was in two thousand. That's about a day for every

0:02:12.080 --> 0:02:15.200
<v Speaker 1>decade of technological and in scientific progress. So that's a lot

0:02:15.240 --> 0:02:20.000
<v Speaker 1>of advancement. And as the weather gets more volatile and extreme,

0:02:20.440 --> 0:02:23.639
<v Speaker 1>will be depending on this weather forecasting capabilities.

0:02:24.000 --> 0:02:26.840
<v Speaker 2>Well, I'm interested in that point. Actually, the increase in

0:02:26.919 --> 0:02:30.240
<v Speaker 2>volatile weather events seems to be something that we're observing

0:02:30.760 --> 0:02:33.640
<v Speaker 2>at pace. How does that affect weather forecasting? Does it

0:02:33.639 --> 0:02:34.519
<v Speaker 2>make it more difficult.

0:02:34.880 --> 0:02:39.360
<v Speaker 1>So weather forecasting is just fundamentally tricky. The atmosphere is complicated,

0:02:39.800 --> 0:02:43.840
<v Speaker 1>it's chaotic, it's uncertain, So every forecast has an element

0:02:43.840 --> 0:02:47.480
<v Speaker 1>of uncertainty. And that's because of incomplete observations, so not

0:02:47.680 --> 0:02:51.040
<v Speaker 1>being able to see exactly everything that's going on in

0:02:51.040 --> 0:02:54.720
<v Speaker 1>the atmosphere, and also just approximations that weather models have

0:02:54.800 --> 0:02:58.359
<v Speaker 1>to make. And now extreme weather makes that slightly trickier

0:02:58.560 --> 0:03:02.240
<v Speaker 1>and even more challenging because extreme weather, I suppose by definition,

0:03:02.520 --> 0:03:05.720
<v Speaker 1>they just have more variation variability, and they've also been

0:03:05.760 --> 0:03:11.000
<v Speaker 1>relatively rare, so that means less data and fewer observations

0:03:11.040 --> 0:03:13.560
<v Speaker 1>over time, and so we just understand these events less

0:03:13.600 --> 0:03:16.519
<v Speaker 1>well and have done less research into them. And then

0:03:16.800 --> 0:03:20.679
<v Speaker 1>these conditions can change really quickly. A hurricane, for example,

0:03:21.400 --> 0:03:26.120
<v Speaker 1>given warmer waters can rapidly intensify, and that quick change

0:03:26.280 --> 0:03:29.320
<v Speaker 1>can catch forecasters off guard. So all of that is

0:03:29.360 --> 0:03:32.560
<v Speaker 1>making it slightly trickier, even though weather forecasting itself has

0:03:32.600 --> 0:03:34.320
<v Speaker 1>never been an easy exercise.

0:03:34.680 --> 0:03:38.240
<v Speaker 2>Well, let's get to artificial intelligence. Then, how are forecasters

0:03:38.360 --> 0:03:41.320
<v Speaker 2>using AI? How can it help this process?

0:03:42.200 --> 0:03:45.600
<v Speaker 1>There are lots of steps to creating a weather forecast,

0:03:46.120 --> 0:03:49.240
<v Speaker 1>and AI can be applied to some of those steps,

0:03:49.360 --> 0:03:51.560
<v Speaker 1>or maybe all of those steps. We can take the

0:03:51.600 --> 0:03:55.000
<v Speaker 1>first step gathering realms and reams of data, because to

0:03:55.360 --> 0:03:57.800
<v Speaker 1>forecast future weather, you need to know what the present

0:03:57.840 --> 0:04:01.480
<v Speaker 1>weather's like, and that requires getting a lot of observations,

0:04:01.560 --> 0:04:05.880
<v Speaker 1>whether that's temperatures or air pressure, humidity and the like.

0:04:06.200 --> 0:04:10.200
<v Speaker 1>So AI can help by helping us gather more data

0:04:10.560 --> 0:04:13.520
<v Speaker 1>from a quantified sense, but also more types of data.

0:04:13.880 --> 0:04:20.440
<v Speaker 1>Whereas current weather models typically are restricted to meteorological observations,

0:04:20.480 --> 0:04:23.520
<v Speaker 1>AI models can now allow us to expand it to

0:04:23.640 --> 0:04:28.240
<v Speaker 1>say maintenance logs for power grids, or even news articles,

0:04:28.560 --> 0:04:32.680
<v Speaker 1>just really gathering more and more information to inform models

0:04:32.680 --> 0:04:37.200
<v Speaker 1>with to then create perhaps hopefully more accurate forecasts.

0:04:37.800 --> 0:04:41.680
<v Speaker 2>How widespread is the use of AI and forecasting currently.

0:04:42.080 --> 0:04:47.240
<v Speaker 1>So a key intergovernmental weather forecasting agency called the European

0:04:47.279 --> 0:04:51.960
<v Speaker 1>Center for Medium Range Weather Forecasting or ECMWF. They've taken

0:04:52.240 --> 0:04:54.960
<v Speaker 1>their AI model into operations earlier this year and they're

0:04:55.000 --> 0:04:59.200
<v Speaker 1>running it side by side with their existing traditional physics

0:04:59.200 --> 0:05:03.240
<v Speaker 1>based the model, and they're soon to launch another iteration

0:05:03.320 --> 0:05:05.960
<v Speaker 1>of that AI model into operations too. So there's lots

0:05:05.960 --> 0:05:09.360
<v Speaker 1>going on in this world. So the ECMWF. They're also

0:05:09.440 --> 0:05:14.400
<v Speaker 1>running AI models from different providers, including China's Huawei the

0:05:14.480 --> 0:05:19.480
<v Speaker 1>tech giant, Google's graph Cast model, Microsoft's Aurora model, and

0:05:19.680 --> 0:05:22.599
<v Speaker 1>China's own Weather Bureau itself is also testing about a

0:05:22.640 --> 0:05:26.479
<v Speaker 1>dozen AI weather models, so a lot of experimentation going on.

0:05:26.560 --> 0:05:31.160
<v Speaker 1>There were probably going to see AI models and traditional

0:05:31.200 --> 0:05:34.200
<v Speaker 1>weather models working in tandem in the years to come,

0:05:34.320 --> 0:05:38.760
<v Speaker 1>rather than seeing AI can completely replace these traditional numerical

0:05:39.040 --> 0:05:40.760
<v Speaker 1>weather model forecasting methods.

0:05:41.240 --> 0:05:43.960
<v Speaker 2>Yeah, I'm curious if there are risks to using AI

0:05:44.120 --> 0:05:47.280
<v Speaker 2>for weather forecasting. Is there things that technology might miss

0:05:47.760 --> 0:05:49.920
<v Speaker 2>if we see a greater implementation of it.

0:05:50.960 --> 0:05:53.880
<v Speaker 1>The way a lot of these AI models are trained

0:05:54.000 --> 0:05:59.560
<v Speaker 1>right now is on historical climate data from decades and

0:05:59.600 --> 0:06:03.039
<v Speaker 1>decades observations in analyzes. But the thing is that the

0:06:03.080 --> 0:06:05.680
<v Speaker 1>past is never perfect predictor of the future, and so

0:06:06.279 --> 0:06:09.960
<v Speaker 1>if we're about to see in a warming climate, more

0:06:10.080 --> 0:06:15.120
<v Speaker 1>extreme weather that previously is not reflected in these historical

0:06:15.200 --> 0:06:17.800
<v Speaker 1>data sets, then there is a risk that AI weather

0:06:17.880 --> 0:06:21.760
<v Speaker 1>models will say underestimates maybe the intensity of a typhoon

0:06:21.839 --> 0:06:25.960
<v Speaker 1>or a tropical cyclone, even as it improves the forecasts

0:06:26.000 --> 0:06:28.520
<v Speaker 1>of a storm's track, for example. So there are some

0:06:28.640 --> 0:06:31.720
<v Speaker 1>risks there, and it's definitely an open and ongoing feel

0:06:31.720 --> 0:06:32.320
<v Speaker 1>of the research.

0:06:32.800 --> 0:06:35.240
<v Speaker 2>So what's the next big development we should be watching

0:06:35.279 --> 0:06:36.360
<v Speaker 2>out for in this area?

0:06:37.360 --> 0:06:40.760
<v Speaker 1>This is less one single big development but rather trend

0:06:40.760 --> 0:06:43.760
<v Speaker 1>and I'd be looking for kind of a shifting distribution

0:06:43.960 --> 0:06:47.520
<v Speaker 1>right of roles in this whole global weather enterprise. Are

0:06:47.520 --> 0:06:52.360
<v Speaker 1>public weather agencies which we've long depended on for weather forecasts,

0:06:52.640 --> 0:06:55.560
<v Speaker 1>are they going to play a slightly different role now

0:06:56.000 --> 0:07:00.640
<v Speaker 1>that private companies, tech firms and also small players increasingly

0:07:00.720 --> 0:07:04.640
<v Speaker 1>jumping into this world of creating and providing weather forecasts

0:07:04.760 --> 0:07:07.159
<v Speaker 1>at a more niche level. So it'll be interesting to

0:07:07.200 --> 0:07:10.600
<v Speaker 1>see how that division of labor between the public forecasters

0:07:10.600 --> 0:07:12.280
<v Speaker 1>and more private players will play out.

0:07:12.960 --> 0:07:15.200
<v Speaker 2>Okay, Mary Hoy, our weather reporter. Thank you very much

0:07:15.240 --> 0:07:18.640
<v Speaker 2>for joining us. For more explanations like this from our

0:07:18.680 --> 0:07:21.679
<v Speaker 2>team of three thousand journalists and analysts around the world,

0:07:21.800 --> 0:07:25.960
<v Speaker 2>go to Bloomberg dot com slash explainers. I'm Stephen Carroll.

0:07:26.080 --> 0:07:28.520
<v Speaker 2>This is here's why. I'll be back next week with more.

0:07:28.720 --> 0:07:29.520
<v Speaker 2>Thanks for listening.