We are using weather data to feed building energy models.

Historically, we would do that with weather from the closest weather station, which can be far away (up to 100 km, I guess).

Recently, we've been investigating the use of ERA5 weather data. This is a fantastic database for us because it provides access to historical data for any given place in the world, for free or close to free.

However, I noticed on a few examples an offset with the weather from the weather station. I won't detail this here.

The point of my question is that I can find ERA5 values but I have no idea of the expected accuracy. In other words, if it says 20°C, does this mean 20°C+-0.2°C, or 20°C+-2°C?

I'd like to tell our engineers how reliable the data I'm providing is.

I guess the accuracy is better for recent data since there are more actual data to fit the model. It may depend on many other factors and vary from one physical parameter (temperature, humidity,...) to another.

Is there any place where I could find such information in a relatively synthetic way?

I've been searching a bit with keywords such as "ERA5" "precision" "accuracy" but found no short answer, if any.

  • $\begingroup$ I've never used ERA5. Do they also include irradiance (e.g. Global Horizontal Irradiance), or just wind/temperature/humidity? $\endgroup$ Jan 26 at 8:21
  • $\begingroup$ @EricDuminil there's much more than this. See CDS. Or Oikolab for easier access to commonly used data. $\endgroup$
    – Jérôme
    Jan 26 at 10:11
  • $\begingroup$ Thanks. I tried to download a few locations over a few time ranges, and didn't manage to get any solar radiation data. $\endgroup$ Jan 26 at 14:24

1 Answer 1


The paper that describes ERA-5 (Hersbach et al., 2020) includes an entire section (Section 7) on the accuracy of the model after data assimilation. As the model assimilates a considerable amount of data from both satellites and in-situ observations, there is always the possibility that differences can occur (e.g., if a specific station was not assimilated, if the difference with the reference fields or other sources of data are large).

The difference in temperature is evaluated in a global sense (their Figure 12) ERA-5 temperature comparison Source: Hersbach et al., 2020.

Time series of monthly and globally averaged ERA5 ensemble spread from 1979 to 2018 at indicated pressure levels for (a) temperature ($K$), (b) zonal wind ($m s^{−1}$), (c) ozone (partial pressure in $mPa$) and (d) specific humidity (in percent of the 1981 to 2010 mean value at the pressure level in question).

  • 2
    $\begingroup$ Thanks! I'll try to dig into the paper next week. If I read the graph correctly, the temperature I'm interested in (surface temperature) should be close to reality more or less 0.4°C, closer for more recent years. It is up to us to decide whether this is better than a weather station that may have its own biases and be a little too far from the considered spot. $\endgroup$
    – Jérôme
    Jan 21 at 17:28
  • 1
    $\begingroup$ I found this page of particular interest: confluence.ecmwf.int/display/CKB/ERA5%3A+uncertainty+estimation. "Uncertainty estimates are available from the Copernicus Climate Change Service (C3S) Climate Data Store (CDS), as part of the ERA5 dataset, and there as 'Product type' = 'Ensemble mean' and 'Ensemble spread'." $\endgroup$
    – Jérôme
    Jan 24 at 16:58
  • $\begingroup$ I downloaded value (reanalysis), ensemble mean and ensemble spread for 2 meters temperature for a given place for 2016-2020. I don't understand why reanalysis value differs from ensemble mean. IIUC, spread is a standard deviation. The spread over those 5 years has a mean of 0.3 (with a std of 0.12) so it should be fair to assume that temperatures value (in this place and date range) have a 0.6°C accuracy with a 0.95% probability (2 sigmas). $\endgroup$
    – Jérôme
    Jan 24 at 17:06
  • $\begingroup$ Sorry, there was a typo in the link. It should work fine and it is the same article you are looking at $\endgroup$
    – arkaia
    Jan 24 at 21:48
  • 1
    $\begingroup$ Answer accepted. I don't think I can get a rough idea of the accuracy as there are many factors, so the best I can get is your link to the paper and the link I added to the page about uncertainty and the API to download uncertainty information. If I have further questions about the values of mean/spread, etc. I'll make them another question. $\endgroup$
    – Jérôme
    Jan 25 at 8:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.