I have been exploring the mean diurnal cycle of different variables of the ERA5 reanalysis in some regions of South America and I always find strange discontinuities or "jumps", as illustrated in the attached pictures. The examples attached are for 15-year region-averaged mean diurnal cycle of 2-meter temperature and precipitation around northeastern Argentina for DJF (hours in UTC in the x-axis and Celsius*10 or mm/day in the y-axis). For temperature you can see an abrupt jump at 10 UTC and other non-smooth transitions at 16 UTC and 22 UTC. For precipitation you can see the same phenomenon at 7 UTC and maybe 20 UTC. I have also found this type of behavior in other regions (even outside of South America) and in other variables such as winds at different levels and wind convergence. The issue in winds is mentioned in the "known issues" of the ERA5 documentation, but nothing about temperature or precipitation. I would not expect this type of behavior in "smooth" variables such as 2 meter temperature and I have never seen something alike in other models or reanalyses.
Hi Julián, I don't have an answer, but I can report being surprised at perhaps similar jumps in diurnal composites of zonal wind at 150 hPa, for example in Fig. 4c,d of Article Diurnal Forcing and Phase Locking of Gravity Waves in the Ma...
. We noticed discontinuities that appear at the same time across wide longitudinal bands, in contrast to a WRF simulation, which has much greater continuity. We attributed it to the potential influence of data assimilation in the reanalysis, but that was speculation. It would be great for someone at EC to comment on this.
Hi. I have also noticed such discontinuities and I wrote to the ERA5 contact mail address. They asked to wait until after the Holidays but they told me they would check this. I'll gladly share their answer.
This is the answer from Paul Berrisford regarding the discontinuities in 2m temperature (which don't appear in radiation).
The ERA5 data on the CDS disks comes from the analyses unless that parameter is only available from the forecasts. The data assimilation system, which produces the analyses including temperature, uses 12 hour windows from 9 - 21 UTC and 21 - 9 UTC, the following day. The analyses should be smoothly varying during each assimilation window but, potentially, there could be discontinuities on going from one analysis window to the next ie from 9-10 UTC and 21-22 UTC. (This agrees with the discontinuities) As you say, we have already documented such cases for u and v. When these were discovered I did check temperature and it was fine. However, that was only for a couple of locations where we had problems with u and v. Therefore, I can't rule out there being problems with temperature at other locations.
The radiation data comes from the forecasts. There are two forecasts per day and in the CDS, forecast steps 1 to 12 are used from the twice daily forecasts to populate the 24 hours of the diurnal cycle. These forecasts begin at 6 and 18 UTC, so there you might expect discontinuities between 6-7 UTC and 18-19 UTC. Fortunately, you don't appear to suffer discontinuities here.
One possible solution to the temperature discontinuities would be to use the forecast temperature instead of the analysed temperature. This data is in MARS and can be accessed using the CDS API, see:
https://confluence.ecmwf.int/display/CKB/How+to+download+ERA5
However, there are disadvantages to using forecast temperatures, namely that the influence of observations is less, only influencing the initial states from which the forecasts are run. For example, the model tends to be biased cold at low levels, so in many regions the forecast temperatures are lower than the analysed ones.
Hi Julián Alberto Giles
I found the same problem in other locations in South America as well, however, they don't show up at the times that Paul Berrisford mention in Tanea Coronato reply. This is an example, in a place near the coast on the NorthEast of Brazil.
I found that the temperature from the levels 1000 and 975 are ok, while the T2m near the surface is not (it jumps at 3, 10, and 15 UTC). Moreover, the land-model output known as ERA5-land, which uses forcing from ERA5, does not show the jumps.
I tried to look at the ERA-Interim dataset as well, but the corse time resolution (6h) of the analysis is not sufficient. I then looked at the forecasts from ERA-Interim, for the runs at 0Z and 12Z, steps 3h, 6h, 9h and 12h. The forecast even more problematic, hence Berrisford's suggestion to use the ERA5 fcts might not solve the issue.
Did you guys find anything beyond what you posted here on Research Gate?
Hello Henrique Barbosa ,
Thank you for your comment. I see you have a 9-10 jump coincident with my results and with the analysis windows Paul Berrisford mentioned. Also you have jumps around 3 and 15, similar to mine, but not mentioned by Paul Berrisford. It is interesting to see that the temperature in different levels does not present this problem, which seems to be also region dependent.
We don't have any additional information. We have not checked the forecast fields yet (I intend to do so). Also, I still don't see any mention to this problem in the ERA5 documentation:
https://confluence.ecmwf.int/display/CKB/ERA5%3A+data+documentation#ERA5:datadocumentation-Knownissues
I'll update this question in the future if I have any news.
Hi Julian,
the jumps are surprising and I don't have much to offer as an explanation, except that I didn't observe such behaviour in ERA5 hourly precipitation dataset over Eastern Himalaya. Can check out figure 8 in the paper for reference. I'm planning to use ERA5 for another analysis over Himalaya and would be very interested in any solutions to your abovementioned problem. Thanks & Regards. Manish
Article Measuring precipitation in Eastern Himalaya: Ground validati...
Hello everyone, I finally had time to check the ERA5 forecast fields. I compared the January 1981 monthly mean 2m-temperature diurnal cycle from the ERA5 reanalysis and the ERA5 forecast fields, for a point in South America. As you can see, the reanalysis has the mentioned discontinuities, but the forecast fields do not. So it is very possible that this is an issue with the data assimilation system, when going from one forecast window to the next.
I do not know of any solutions to this, other than using the forecast fields instead of the analyzed fields. An alternative would be to use the fields from ERA5-Land, which apparently do not have this problem according to Henrique Barbosa 's example. This issue is still not mentioned in the ERA5 "know issues" documentation. I have sent a new email to the ECMWF support contact, I will update you with any news.
Hello everyone, I have now received a new answer from the ERA5 research team. The issue is indeed related to the initialization of the analysis windows. They have updated the documentation and added the case for temperature and humidity discontinuities in the "Known issues" section:
https://confluence.ecmwf.int/display/CKB/ERA5%3A+data+documentation#ERA5:datadocumentation-Knownissues
They are working on a more in-depth analysis of this issue that will be published in the future.
A final but important point: they confirmed this issue cannot be fixed. Using the forecast fields is a possible workaround, however, beware that the forecast near surface temperature in ERA5 tends to have a cold bias over most parts of the globe.
- Badges
- Science topic
- Similar topics
- Geoscience
- Atmospheric Sciences
- Climate Science
- Climate Change