ERA5 Analysis of Cold

During the interior cold snap of the last week of 2019, when the temperature dropped well below -50°F in many places, I was eager to see what the new state-of-the-art ERA5 reanalysis would say about the minimum temperatures.  I've used the ERA5 data on previous occasions (e.g. see here), and will continue to do so, because it's an exciting new product from Europe's weather and forecasting science powerhouse, ECMWF.  Besides this, ERA5 is becoming widely used as essentially ground truth data across the globe, so it's of great interest to see how it performs in extreme or unusual weather situations.

First, here are some maps of the ERA5 daily minimum temperatures during the cold spell.  Note that I've used a 24-hour period from 3pm to 3pm AKST (midnight UTC) for simplicity in processing the data.  Click to enlarge.

When comparing to the December 27 satellite-estimated values that I showed in a previous post, it's clear that the area affected by -50°F or lower is considerably smaller in the ERA5 data, especially to the south of the Yukon River.

The absolute minimum on the ERA5 31-km grid is -58.7°F, and this is found right over Allakaket on December 27; this temperature matches very well at Allakaket itself, but as discussed before, the satellite suggests it was a bit colder on the Kanuti Flats.  Moreover, the ERA5 analysis is nowhere near as cold in the vicinity of the Eureka coop (MLYA2) report of -65°F, although of course we can't expect a 31-km resolution model to capture localized valley cold pools - we would need a far finer grid to hope for success in that situation.

It's interesting to see that the model does have a localized area of cold near the Nowitna CRN site (NWTA2), and so I thought it would be worthwhile to do a more detailed comparison of ERA5 vs CRN temperatures.  Of course the quality of the CRN measurements is extremely high.

First, the big picture: all daily mean temperatures since March 2017, when the Nowitna (Ruby 44 ESE) CRN site first started working reliably (after being installed in 2014).  See below - the comparison is fairly good at temperatures above freezing, but there are evidently some problems in the cold season; in particular, ERA5 tends to have a high bias in cold conditions, and it's large in some cases.

Here's a chart of daily mean temperatures this winter since October 1.

The main problem is immediately obvious: ERA5 stays much too warm when the temperature drops off quickly at the CRN site.  Each of these cold snaps no doubt corresponds to a period of clear and calm weather that allows a strong temperature inversion to develop, and evidently the model's boundary layer physics do not handle this situation well.  It's a bit disappointing, considering the pedigree of the ERA5/ECMWF models, but we must admit that it's a tall order to reliably represent such shallow layers of intense temperature inversion in any model.

Here are charts from the past two winters.  The same problem is evident, although not to the extent of this winter.  Interesting the mean bias from October-March was exactly the same in both 2017-2018 and 2018-2019 (ERA5 3.1°F too warm), but the bias is running at +4.9°F so far this winter.

In summary, ERA5 clearly has some problems with extreme cold in Alaska in winter.  In the recent interior cold snap, the extent of coldest conditions was too small, although part of this is certainly related to the grid-box averaging that is intrinsic to the model; the model simply doesn't represent valley-floor locations when there is higher terrain elsewhere in the grid cell.  Looking at the Nowitna CRN, there's a clear tendency for a warm bias that shows up during cold spells, and this seems to be a symptom of inadequate surface layer physics as well as insufficient horizontal resolution.

All of this goes to show that we still can't overestimate the value of real ground-truth observations such as those provided by the wonderful CRN network.  And as an aside, I for one am very grateful for the efforts of the NWS personnel who finally got the Nowitna site working properly - it's a great location and we're lucky to have such a treasure trove of data.

ERA5 Data for Alaska - Including Download Link

Back in November I took an initial look at the new ERA5 reanalysis data set from the world-leading ECMWF weather modeling and forecasting center in Europe.  At the time, the ERA5 data was only available for 2000-2017, but the reanalysis now extends back to 1979, and a further extension to 1950 will soon come online.  The high quality of the data assimilation and modeling framework that's used to produce the reanalysis makes this a real treasure trove of historical climate data.

It's an interesting exercise to compare the ERA5 data for Alaska to NOAA's climate division data, produced by NCEI.  For many years the climate division data was only available for the lower 48, but in 2015 the data set was expanded to include 13 climate zones in Alaska; here's a map.

To facilitate a direct comparison, I calculated area-averages for several ERA5 variables within each of Alaska's climate divisions.  For example, there are 605 ERA5 grid cells that at least partially intersect the Southeast Interior division; so I calculated the area of the intersection for each grid cell and added up the fractional contributions to the total area of the Southeast Interior zone.

Here's a chart showing the mean temperatures for January and for July in the Southeast Interior (which includes Fairbanks).  Aside from a modest cold bias in the ERA5 values in January, the performance is outstanding.

The situation is not quite as good in the North Slope division, which is not surprising as the observing network is more sparse, and moreover weather analysis and forecasting models (like the ECMWF model that underpins ERA5) often have a more difficult time with atmospheric physics in the Arctic.

Interestingly the 1979-2018 linear temperature trends are similar for January, but the ERA5 trend is much smaller than NCEI's trend for July in the North Slope division.

Looking at precipitation, ERA5 does fairly well for the Southeast Interior in both January and July, but again the agreement is not as good for the North Slope.  Precipitation is always a major challenge for reanalysis, and so these results are pretty good.

Finally, I did a quick comparison of ERA5 solar radiation to the CERES gridded data for the Southeast Interior, and again I used an area average for both data sets.  The results show a very close correspondence for the month of March, but there is only modest agreement in July.

We could of course keep going with all sorts of comparisons between ERA5 and other data sets, and between ERA5 and historical climate observations around the state, but there's no doubt that ERA5 is a very high quality reanalysis.  Beyond the pure fidelity of the data, however, the real value of the reanalysis is that it's spatially and temporally complete; and ERA5 even includes uncertainty estimates, although I haven't looked at that aspect yet.

For readers who might like to take a look at the Alaska data themselves, the following link provides the area-averaged data for the 13 climate divisions, including mean temperature, precipitation, solar radiation, and 10m wind speed.

https://drive.google.com/drive/folders/15RUOjHKVmxFJnnW1a9ZaMpQRwR-G6yPD

New ERA5 Reanalysis

One of the more exciting developments in the world of weather and climate science this year has been the release of new data sets by the European Union's Copernicus Climate Change Service.  The program is funding the free and open distribution of vast quantities of data through the Climate Data Store, so there is almost unlimited scope for new research as well as commercial development using the data.

The data set that I'm most excited about is the latest generation of reanalysis from the European Centre for Medium-Range Weather Forecasts (ECMWF), which is well-known for having the most accurate global weather forecast model in the short-medium-range time frame (out to two weeks in the future).  I've often used reanalysis data from NOAA on this blog, and indeed NOAA's global reanalysis from 1948-present is heavily used worldwide and is extremely valuable.  However, the NOAA reanalysis relies on a model that is very out of date now.  Happily, the new ECMWF reanalysis - using the ECMWF's top-notch modeling capability - is now coming online via the Copernicus program; the data are currently available back to 2000, but next year we'll see the product extended back to 1950.

Here's an article about ECMWF's new ERA5 reanalysis:

https://www.ecmwf.int/en/about/media-centre/science-blog/2017/era5-new-reanalysis-weather-and-climate-data

Back in 2015 I did a brief comparison of NOAA's reanalysis data with real observations from Fairbanks; here's one of the figures, showing the very poor correlation of reanalysis to actual temperature and precipitation in summer.

The chart below is a similar figure using ERA5 data for the nearest grid point to Fairbanks, which happens to be located just to the south across the Tanana River (the grid spacing is about 20km).  The performance is impressive.  Now admittedly the correlations ought to be very high for temperature, because the ECMWF model uses surface observations to refine its gridded estimates of evolving weather conditions hour by hour.  However, precipitation is predicted by the model over short time intervals, so the model does not "know" how much precipitation occurred in reality; and neither ground-truth data nor radar estimates are used to improve the estimates.  Given that the ERA5 precipitation data is purely a (short-range) forecast, I think it's very impressive that the monthly correlations are as high as ~0.8 in May through July, when hard-to-predict showers and thunderstorms produce most of the rain.

Here's a look at correlations of daily rather than monthly temperatures through the year.  Daily low temperatures are generally more difficult to get right than high temperatures, because the warmest conditions of the day tend to be more closely tied to the more homogeneous, well-predicted temperatures of the free atmosphere above.

Finally, the wind speed estimates from the model are not as impressive; apparently the low-level wind regimes near Fairbanks are a challenge even for the world's best global modeling system.

In due course I will be acquiring a larger volume of the ERA5 data and will have a chance to do a more extensive analysis; and it would be fun to set up an online map catalog of ERA5 data for the Alaska domain.  If anyone has an interest in helping out with such a project, let me know - perhaps there could be a collaboration.