AI Weather Forecasts

I'm sure many readers have come across recent headlines about new breakthroughs in Artificial Intelligence for weather forecasting; here's one example:

https://arstechnica.com/science/2023/11/ai-outperforms-conventional-weather-forecasting-for-the-first-time-google-study/

It's not always easy to distinguish between hype and reality when it comes to claims of new technological advances.  Is the enthusiasm justified in this case?  Are weather forecasts about to see a revolutionary step forward in accuracy?  I'll offer a few words of my own perspective, as someone in the "weather business".

First, there's no doubt that AI and Machine Learning (closely related concepts) have made amazing strides for weather prediction in the past few years.  In early 2021, ECMWF (the world's leading "traditional" weather forecasting organization) published a "road map" for the future of AI weather forecasts, anticipating a rather gradual pace of innovation and change; but it turned out that several leading technology companies achieved remarkable success in just the next 2 years.  Crucially, the latest AI models were suddenly revealed as being able to match or even beat the ECMWF's forecast accuracy according to some metrics.  ECMWF wrote about it here:

https://www.ecmwf.int/en/about/media-centre/science-blog/2023/rise-machine-learning-weather-forecasting

The headline result - exceeding the ECMWF's basic skill level - is a big deal.  It means these new models are legitimate competitors, and it's truly remarkable that the AI scientists have achieved this so quickly, with relatively early-stage, experimental models.  In contrast, the accuracy of traditional forecast models has developed steadily but very slowly over the years, relying on bigger and faster computers, more satellite data, and incrementally better physics in the models.  Modern weather forecast accuracy is a great scientific accomplishment; but now without warning it is being equaled (by some metrics) with an immature technology that presumably has a lot of room to improve.

Here's an example figure from ECMWF showing gradual improvement of skill for predicting Northern Hemisphere 500mb height at various lead times.  After all this work, the idea that an upstart new technology can suddenly jump in with similar or better skill is quite surprising and perhaps difficult to swallow!

But notice that I alluded to specific metrics that show the new models in a favorable light, and this is because the first generation of AI models has significant limitations.  For example, it seems the models don't do particularly well with extreme events, because they are guided by (constrained by) the historical data rather than the laws of physics.  It seems to me that models trained purely on historical data will always struggle in this way, but hybrid statistical-dynamical models are an obvious extension that would be more likely to handle unprecedented events.

Another area where the new models aren't yet fully capable is in terms of handling uncertainty.  Forecast centers like ECMWF have become adept at running traditional models in a way that encapsulates and predicts uncertainty, and this is extremely important for valuable real-world forecasting: we need to know the plausible range of possible outcomes.  In contrast, the first-generation AI models just give a single answer to the question: "given today's weather, what will the weather be N days from now?"

It's also worth noting that - at least for now - the new AI models rely on the traditional models to provide the initial conditions for the forecast, i.e. to specify what "today's weather" is in great detail around the globe.  This "initialization" process is itself a tremendous scientific achievement, requiring a very advanced model to "assimilate" data from the entire observing system and create a single best guess of what's happening at every location.

What Difference Does It Make?

It's worth asking what practical difference there will be for weather forecasts created with AI models rather than traditional models.  This is of course difficult to foresee, but I suggest the answer is "not much" until the mainstream weather industry fully comes to grip with the new technology and builds it into the forecasting process.  The AI models have been developed by big technology companies (e.g. Google and NVIDIA) that are not (yet) in the business of selling forecasts, and the vast majority of meteorologists and meteorological scientists work elsewhere in government, academia, and traditional private sector weather companies.  These two "worlds" will need to come together if the new technology is to be deployed widely, and the process won't be quick or painless.

In terms of specific predictions, I can see two things happening.  First, the big-tech AI models will be licensed to private weather companies who find value in the AI forecasts and can build products for customers.  The leading forecast centers like ECMWF will also build their own AI models and provide the results to users, and so the industry will gradually adapt to the new source of information; but the traditional methods certainly won't be ditched any time soon.

Second, I think the AI methods will be used to develop much better forecasts for some specific high-impact problems that are not handled well by traditional models.  An example might be fire weather forecasts: for example, an AI model trained on past weather-driven fire events could provide powerful guidance for future risk.  The deadly fires in Hawaii and California in recent years might have been much better predicted with a specific application of the new technology, allowing aggressive early evacuations in those rare and dangerous weather situations.

I'd be glad to hear comments and insights from readers.  What else do AI methods have to offer that physics-driven "deterministic" models can't provide?  What other forecast problems are poorly handled by today's usual weather guidance and might be particularly amenable to historical/statistical methods like AI?

Wild Weather in the South

Weather is never very far from the news in a place as big as Alaska, and the more extreme events aren't usually good news.  Yesterday a major Gulf of Alaska storm barreled into a cold air mass and high pressure over southern Alaska, and the results were dramatic: damaging winds in the Matanuska Valley, and a major landslide down south in Wrangell.  Sadly the landslide brought loss of life, and perhaps worse than the 2020 Haines landslide.

ADN news links:

https://www.adn.com/alaska-news/mat-su/2023/11/21/most-power-restored-and-roads-cleared-following-mat-su-windstorm/

https://www.adn.com/alaska-news/2023/11/21/large-landslide-south-of-wrangell-closes-highway-draws-search-and-rescue-team/

The rain in the Southeast doesn't appear to have been particularly unusual for the region; the map below shows 72-hour totals (click to enlarge).  But it was pretty wet earlier in the autumn, so there may have been a compounding risk.  Sitka is in the top 10 for rainfall since September 1.

As for the winds around Palmer and Wasilla, the map below shows how localized the wind storm was in the greater Anchorage area (showing yesterday's peak gusts in mph).

This looks like a great example of channeled downsloping flow from the east, driven by a strong pressure gradient.  The map below shows the sea-level pressure analysis at 3pm yesterday, courtesy of Environment Canada.  Check out that pressure gradient over the southeastern coast of Alaska; high winds were seen throughout the region.

A Few Notes

A few comments on various topics.  First, the snow onslaught in Anchorage: the recent 10-day total of 38" amounts to the 3rd highest on record.  Last year's December snow was even a bit greater, and the top event was back in February 1996.

It's really remarkable to see such a large amount of snow at the airport - where there is typically less snow than across most of the area - and in back-to-back years.  With La Niña last winter and El Niño this winter, we can't pin the blame on similar climate forcing, so perhaps it's just "luck of the draw".

Second, the Tanana River at Nenana froze up last week on Wednesday the 8th.  The scene is suitably wintry today:

But the mighty Yukon River hasn't frozen over at Dawson yet; here's the latest video from the webcam:

The situation at Dawson looks about the same as last year at this time:

https://ak-wx.blogspot.com/2022/11/historical-context-for-warmth-freeze-up.html

Long-time readers will recall discussion of the lack of a proper freeze-up all winter long at Dawson in some recent years:

https://ak-wx.blogspot.com/2020/02/yukon-frozen-at-dawson.html

This autumn has certainly been on the warm side of normal in Dawson, but it's not breaking records for lack of freezing potential: here's a chart of accumulated freezing degree days (accumulated temperature differential below 32°F):

This is from the reliably-reporting "LRP" site in Dawson, not the airport, where many observations are missing.  The city does have a much longer history of climate observations, of course, but it seems there are some discrepancies between the different climate sites that would affect any long-term trend analysis.

And finally, I'll highlight the remarkable lack of sea ice along Alaska's west coast at this time, related to persistent southerly winds and warmth.  Rick Thoman posted a graphic to illustrate:

"Sea ice concentration analysis for Wednesday and the same date last year from the National Weather Service Alaska Region Sea Ice Program. Much less ice in the southern Chukchi and Bering Seas currently than this date last year. November 15 and effectively no ice in Kotzebue Sound is especially shocking, even by recent norms."

https://mastodon.social/@AlaskaWx@alaskan.social/111417939896937766

Here's a longer archive of November 15 analyses, from 2021 (top) to 2018 (bottom):

South Coast Snowstorm

Yesterday a number of 747 cargo planes were spotted at Fairbanks airport, and that usually means only one thing: Anchorage is having a big snow event.  And indeed they were: 2-day snowfall amounts ranged from 17" at the airport to 30" on the Anchorage Hillside and over 36" up Turnagain Arm.  Here's a graphic courtesy of NWS Anchorage, showing storm totals as of this morning:

With substantial snow earlier in the week as well as more today, the weekly total is really something: Brian Brettschneider indicates that the airport's 6-day total of over 28" is the third highest on record in the Anchorage climate history.  The only larger snow events in that time frame over the last century or so were in February 1996 and December 1955.  Quite a way to start off the winter!

The snow depth of 21" yesterday afternoon is easily the greatest for this early in the season in Anchorage.

Of course, far more snow fell in topographically favored locations.  Courtesy of NWS Anchorage (and publicized by Brian), here's an astonishing report from the Richardson Highway:

The liquid-equivalent precipitation estimates from the NWS are wild too, with over 10" estimated in the high terrain on Wednesday, and another few inches yesterday.

Here's a satellite loop showing the powerful cyclone and associated "atmospheric river" that brought such prodigious moisture to the south coast.

Yesterday morning's 3am surface analysis shows the 964mb low tracking directly into the Kenai Peninsula and Anchorage area.  Winds gusted as high as 60mph in Yakutat.

And here's a nice loop of the Anchorage-area radar reflectivity, courtesy of Randy Chase on Twitter:

https://twitter.com/DopplerChase/status/1723058354732282352

October Climate Data

Rick Thoman has a nice write-up of October's Arctic and Alaska climate anomalies here:

https://alaskaclimate.substack.com/p/october-2023-climate-summary

The precipitation contrasts across Alaska were striking: it was much drier than normal for the time of year in the southwest and south-central, but the North Slope and eastern Alaska were notably damp.  Rick notes that the Utqiaġvik CRN site measured 1.68" of liquid-equivalent for the month - the highest on record there for October (at either the CRN site or the much longer Barrow/Utqiaġvik climate site).  It's also the fourth consecutive month with more than an inch of precipitation at the CRN location.

Snowfall was also relatively abundant in eastern Alaska: as high as 19" in Tok and 17" in Eagle.  Also 18" on Keystone Ridge outside of Fairbanks.

Here's a look at snow water on the ground at 5-day intervals through the month, expressed in terms of historical percentile, and estimated by the ERA5 model:

The snowpack was established on October 6 in Fairbanks (considerably earlier than normal), but very little snow remained by the end of the month in the southwestern quadrant of the state.  The last week of the month was very mild, but chilly weather earlier in the month allowed for a near-normal monthly average temperature for much of the interior.  The Aleutians, North Slope, and Southeast were notably warmer than normal, however.

As for wind, October was relatively calm for the southwest and south-central, but windier than normal in the north and northeast.

Here's the mid-atmosphere pressure pattern that contributed to the surface climate anomalies: the ridge over southwestern Alaska explains the dry and calm weather there.

Sea surface temperatures remain much higher than normal in the western Bering Sea and northwestern North Pacific, so we can expect that winds from that direction will tend to bring unusually mild conditions to Alaska in the coming months.

Given that we have a strong El Niño in play this winter, which tends to produce a strong Aleutian low and frequent southwesterly flow across Alaska, I'd say the chances of a very mild winter are much higher than normal.  NOAA's Climate Prediction Center agrees:

For a comprehensive discussion of the winter outlook, check out Rick Thoman's November 17 presentation here:

https://uaf-accap.org/event/nov-climate-outlook/

Turning Warm

After relatively cool conditions for much of autumn in Alaska as a whole, the last week of October brought a change to unusual warmth.  Blame a big ridge extending north from the Gulf of Alaska, allowing warm air to be transported northward into the western part of the state and up to the Arctic.  Here are the average 500mb height and 850mb winds for the week ending October 30:

And the resulting 850mb temperature departure from normal:

Warmth in Alaska's far north has been an oft-repeated theme in the past two decades, and this year Utqiaġvik ended up with a top-5 warm October, although it wasn't as extreme as 2016 and 2019.  Here's the now (I trust) well-known chart of October temperatures there, showing the profound change in the local climate from loss of October sea ice.

For the state as a whole, October 30th was the warmest day of the month in comparison to normal.  The figure below shows the widespread distribution of unusual warmth, with the statewide average climbing into the 95th percentile for the time of year.  The standardized anomaly numbers reflect my calculations that account for seasonal skewness, as discussed recently here.

Radar Trends

Back in early August, when interior Alaska was seeing an unusually late spell of thunderstorm activity, I commented that it would be interesting to work with the Fairbanks radar data to investigate the variability of strong storms:

https://ak-wx.blogspot.com/2023/08/severe-thunderstorm-warnings.html

I ended up digging into this, and I've succeeded in producing an archive of integrated storm activity back to 2014.  The Fairbanks (Pedro Dome) radar has been around for a lot longer than that of course, and there is some historical data available back to 1997, but unfortunately there is just too much missing data in the early years to use it in this analysis.  There's also a complete gap from 2004 through 2013, and regrettably NOAA/NCEI says that data does not exist any more.

But we'll work with what we have.  My approach was to calculate the spatial area for which the radar reflectivity exceeded various thresholds at each radar scan, across the entire available scan domain, as archived in NOAA's Big Data project.  I used the base (lowest scan elevation) reflectivity and extracted the data on a 1km grid.

There is a myriad of ways that the results could be analyzed, but for starters I took the May-September average of each day's maximum area at or above the reflectivity threshold - such as 40 dBZ, see below.  A 40 dBZ reflectivity is typical of moderate convective rainfall in summer; a typical robust thunderstorm core would be more like 50 dBZ.

Of course many days have no thunderstorms and little or no rainfall across the domain, and most of each season's total activity is concentrated in a rather small number of days, but the season-long average of daily peak activity seems to provide a suitable integrated measure.  Note that I excluded days with less than 100 reflectivity scans in 24 hours; about two-thirds of days have more than 200 scans in May-September.

The overall year-to-year variability in 40+ dBZ radar echoes does not seem particularly great, but the monthly breakdown shows that both July and August of the last two summers were considerably more active than the decadal average.

Raising the bar to a 50 dBZ threshold emphasizes the high level of storm activity in the past two summers, especially in July (both years) and August (this year).  August 2023 actually stands out quite dramatically compared to the previous 9 years:

Taking it up still another notch to 60 dBZ, which would be typical of a strong storm producing at least some hail (possibly large), summer 2023 stands out even more compared to prior years, and there's a strong suggestion of an upward trend.  Despite the very short period of record, the linear trend in the first chart below is highly statistically significant.

This data obviously begs for further analysis and investigation, and I'll aim to do that as time permits.  For now, it's just interesting to note that perhaps the unusual spate of severe thunderstorm warnings issued by NWS Fairbanks this summer wasn't as out-of-line as I thought at the time.

Kaktovik Wind

Blizzard season is off to an early start in Alaska's capital for that kind of thing: Kaktovik, on Alaska's eastern Arctic coast.  Snow has been flying since Thursday morning, and winds increased to 50 mph yesterday evening (with gusts to 60 mph).  Thanks to reader Mike for bringing this to my attention.

October is rather early for this kind of weather in Kaktovik, and it's been more than 10 years since such high winds were observed in October.  In winter proper it's more common, with more than 5% of days seeing sustained winds over 45 mph in November through February.

Here's the meteorological setup: high pressure over western and central Alaska, and a deep low pressure system over Banks Island (as of 4pm AKST yesterday).  My clumsy annotation shows the direction of the strong pressure gradient, and Barter Island's location is indicated with the red dot: click to enlarge.

Statewide Temperature Index

Last week I mentioned the UAF/ACCAP statewide temperature index (developed by Rick Thoman and Brian Brettschneider), which is a useful tool for quickly visualizing the overall ebb and flow of temperature anomalies in Alaska.  Here's the last year or so (click to enlarge):

Two of the more striking anomalies in recent months were the extraordinary warmth of late summer and the April cold spell.

While I'm a big fan of the index that Rick and Brian came up with, I've been tinkering to see if it can be improved still further.  One of the limitations of the existing index is that it assumes a Normal distribution for daily mean temperatures, but we know that the distribution is significantly skewed at certain locations and times of year.  Here's an old post illustrating this:

https://ak-wx.blogspot.com/2014/03/temperature-skewness-maps.html

To account for skewness, I was able to fit a Skew-Normal distribution to the historical temperatures at each of the 25 ASOS locations used in the ACCAP index.  Using these distributions - which vary through the year - the statewide index can then be recalculated from the newly standardized anomalies at the individual locations.

This modification makes a fairly big difference at certain times of the year.  An extreme example is late March 2019, when extremely warm weather occurred at a time of year when skewness is negative for most of the state.  In other words, negative departures from normal tend to be larger than positive departures, and so a large warm anomaly is actually much more unusual than a Normal distribution would imply.  Here's an illustration of the standardized anomalies on March 22 that year, using a Normal distribution (top) and a Skew-Normal distribution (bottom):

The standardized anomalies are much larger in the Skew-Normal framework, and the statewide anomaly jumps from +2.15 to +2.71.  In terms of degree-of-rarity, the Normal distribution estimates a 98.4% percentile for the day, but the Skew-Normal distribution shows 99.7%.  Both are in the upper tail, but accounting for the negative skewness reveals that the warmth was statistically much more extreme.

It works the other way too: a few months later in 2019, extreme heat occurred in early July, but with positive skewness around Alaska's coastal regions, the large warm anomalies were somewhat less significant than a Normal distribution would indicate.  The statewide standardized anomaly drops from +3.16 to +2.88, which although not a large change implies more than a 50% increase in the climatological probability of such an event (0.08% vs 0.20%).

Of course we could find examples on the cold side too.

In addition to the skewness modification, I also adjusted the statewide index calculation to account for the fact that the (weighted) average of the 25 individual (standardized) anomalies has larger variance in winter than in summer.  This is because there is more spatial correlation of temperatures across the state in winter, so the 25 (standardized) temperature measurements are more prone to move up and down together, and the combined average has a relatively high variance.  In contrast, summer temperature anomalies are more prone to cancel each other out across the state in summer, and the statewide average (of standardized anomalies) varies less.  (This is a completely separate issue from the reduced variance of actual temperatures in summer.)

So rather than using a single number to convert the statewide mean into its own standardized index, I used a seasonally varying number, and this varies by nearly 40% between about January 20 and June 20.  The result is that my modified index shows systematically smaller statewide anomalies in winter than the ACCAP method, but larger anomalies in summer.

Here's a comparison of the past year's statewide standardized anomaly from the two methods.  Notice that the revised method shows smaller values in winter, larger values in summer, and significantly less extreme cold in April (because of the negative skewness).

And here's the actual ACCAP index, which shows the statewide value in terms of percentile/frequency rather than standardized anomaly.

The major difference between the anomaly view and the percentile view is that the percentile view doesn't differentiate the magnitude of extreme events, i.e. once you're in the top or bottom 5-10% of events, they all look quite similar in the percentile view.  In contrast, the anomalies clearly delineate the amplitude of extreme events, but on the other hand they don't provide an intuitive measure of frequency: most people don't have a sense of how rare a 2 or 3 standard deviation event is.  Both perspectives are useful.

My next step with this work is to set up an automated process to create daily graphics and perhaps a simple webpage for monitoring purposes.