...

Seasonal forecasts and the Copernicus Climate Change Service (C3S)

Introduction to seasonal forecasting

The production of seasonal forecasts, also known as seasonal climate forecasts, have experienced has undergone a huge transformation in the last few decades: from a purely academic and research exercise in the early '90s to the current situation where several meteorological forecast services, throughout the world, conduct routine operational seasonal forecasting activities. Such activities are devoted to provide providing estimates of statistics of weather in the seasonal and monthly on monthly and seasonal time scales, and they lie in a place which places them somewhere between conventional weather forecasts and climate predictions.

In that sense, even though seasonal forecasts share some method methods and tools with weather forecasting, they are part of a different paradigm which requires treating them in a different waysway. Seasonal forecasts instead Instead of trying to answer to the question "how is the weather going to look like on a particular location in an specific day?", they seasonal forecasts will tell us how likely it is that the coming season will be wetter, drier, warmer or colder than the 'usual conditions ' for that time of year. This kind of long term predictions are feasible due to the behaviour of some of the earth Earth system components which evolve more slowly than the atmosphere (e.g. the ocean, the cryosphere) and in a predictable fashion, so their influence in on the earth system atmosphere can add a noticeable signal.We consider here as seasonal forecasts data both the higher-frequency (sub-daily and daily) data outputs from numerical earth system weather or climate models and the derived monthly and seasonal products, up to a few months ahead of their initialization date.

Seasonal forecasting within the C3S

The C3S seasonal forecasting forecast products are based on data from several state-of-the-art seasonal prediction systems. Multi-system combinations, as well as predictions from the individual component participating systems, are available. The centres currently providing forecasts to C3S are: ECMWF, The Met Office and Météo-France; in the coming months data produced by , Deutscher Wetterdienst and (DWD), Centro Euro-Mediterraneo sui Cambiamenti Climatici will be included in the C3S multi-system(CMCC), National Centers for Environmental Prediction (NCEP), Japan Meteorological Agency (JMA) and Environment and Climate Change Canada (ECCC), and you can find the details about when each system was introduced in the information contained in the Summary of available data.

Each model simulates the Earth system processes that influence weather patterns in slightly different ways, makes slightly different approximations, leading to different kinds of model error. The These errors typically increase with the increase of integration time, so that the accumulated model errors become significant in comparison to the signal that the model is meant to be predictedpredict. Some of those such errors are shared by the different models but others are not, so combining the output from a number of models enables a more realistic representation of the uncertainties due to model error. The current scientific knowledge have shown that, in In most cases, such combined forecasts are, on average, more skilful than forecasts from the best of the individual models.

Currently, there are graphical products available in the C3S seasonal service offers graphical forecast products, available on the C3S web site, and public access to the forecast data is provided , via the C3S Climate Data Store (CDS). 

Seasonal forecasts are not weather forecasts: the role of the hindcasts

Due to its long leadtimes, a few months since the start date of the forecast, some systematic errors appear that can completely kill the signals that are expected to be predicted. To avoid that, seasonal forecasting systems work with a reference climate to determine how predicted values differ from what is normal for a given region and time of yearSeasonal forecasts are started from an observed state of (all components of) the climate system, which is then evolved in time over a period of a few months. Errors present at the start of the forecast (due to the imprecise measurement of the initial conditions and the approximations assumed in the formulation of the models) persist or, more often, grow through the model integration, reaching magnitudes comparable to that of the predictable signals.  Some such errors are random; the effect of these on the outcome is quantified through the use of ensembles. Some errors, however, are systematic; if these systematic errors were determined, corrections could be applied to the forecasts to extract the useful information. This is achieved by comparing retrospective forecasts (reforecasts or hindcasts) with observations. The same forecast system is run for past dates, so that kind of model climate can be estimated, and the forecasts be bias-corrected with respect to that model climate. Those forecast runs for earlier years are known as 'hindcasts' or 're-forecasts' and they are a key concept in seasonal forecasting, in such a way that a forecast itself it's not useful without relating it with the relevant hindcasts.several starting points in the past in the same way as a forecast would be run (with only knowledge of the starting point), for the same length of time as an equivalent forecast. The resulting data set constitutes a 'climate' of the model, which can then be compared with the observed climate of the real world. The systematic differences between the model and the real world - usually referred to as biases - are thus quantified and used as the basis for corrections which can be applied to future, real-time forecasts.  Given the relative magnitude of such biases, some basic corrections are essential to convert the data into forecast information - therefore a forecast by itself is not useful without relating it to the relevant hindcasts.

The image below is In the image shown below as an example of the crucial role of the hindcasts in seasonal forecasting, a . It shows the time evolution of monthly-average of temperature in a given region is plotted. All the forecast ensemble members (green lines) seem to point to colder than usual conditions when directly compared to the observed climate (blue line). If the model climate derived from the hindcasts (red line) is taken into account the conclusion is quite different, with all the forecast ensemble members now consistently pointing to warmer than usual conditions.

Image Removed

, between May and December: the blue line represents the average conditions observed over a period in the past (the reference period; in this case, 1993-2014), the red line is the equivalent model climate average over the same reference period. The difference between the two clearly indicates a significant cold bias, increasing with the time into the forecast. An ensemble of forecasts for a particular year is shown as green lines. When compared to the observed 'normal', all green ensemble predictions are for colder-than-normal conditions - not necessarily surprising when remembering that the model is systematically colder than the real world. However, when comparison is made with the model's 'normal' (the red line), all forecast ensemble members indicate warmer-than-normal conditions. Clearly, as the forecast is an output of the model, the latter comparison is the more appropriate.

Image Added

As well as playing an essential role in the correction of Additionally to their unavoidable role to correct those systematic errors, hindcasts are also used to assess the skill of the different seasonal forecasting systems. In that way, the real-time forecasts are expected to behave with the same skill that the system has shown for past dates, i.e. the skill shown by the hindcastsforecast systems (by comparing each of the forecasts for the years in the reference period with the respective observed conditions). Information on forecast skill is important to avoid overconfident decision making.

Note that for reasons related with the availability of computing resources, the hindcasts usually have fewer ensemble members per start date than the real-time forecasts, e.g. ECMWF SEAS5 5 has 51 members for the real-time forecasts, and just 25 members for the hindcasts.

Seasonal forecasting systems' versions and updates

Every forecasting system that contributes to C3S will have a different lifetime, so different versions of the systems are expected to be changed, upgraded from by their original institutions. For the real-time forecasts just one version of each one of the contributors will be made available to C3S as at a real-given time forecasts. For instance, in November 2017 ECMWF has changed its operational seasonal forecasting forecast system from system 4 to SEAS5, but both systems will be were kept routinely running in parallel at ECMWF for a while. Despite that factHowever, the only version of ECMWF seasonal forecasts available at C3S from November 2017 onwards will be is from SEAS5.

How

...

do seasonal forecasting systems build their ensembles? And how are data

...

produced?

"Burst" vs. "lagged" mode

In the last few decades, in the earth system modelling it has been an established technique Earth system prediction has established the use of "ensemble" runs to take into account , to quantify the effect of errors due to both the uncertainty in the initial conditions and model deficiencies. This means that the forecasting systems produced a set of "slightly" different runs of the same forecast which form - the members of the ensemble , in a way that the outcome - and thus the output of the forecasting forecast system is not a single model output solution, but a set of different results which allow to produce a forecast in terms of a probability distribution as opposed to a single deterministic forecastsolutions. Since, by design, all ensemble members are equally likely, the forecast offers a distribution of outcomes, rather than a single deterministic answer.

Different techniques are commonly used to build the different members of an ensemble forecast, but one of the most common ways to create a set of slightly different members that mainly maps so that they sample the uncertainty in the initial conditions, is the use of a "lagged" approach in the start dates.:

• "Burst" mode: all the members are initialized (with conditions on the same start date) at the same value , but from slightly different (perturbed) initial states, intended to sample the uncertainty in observations. (e.g. all members initialized on 1st March 2017, ; this is the case for ECMWF's system)
• "Lagged" mode: members are initialized in on different start dates, the differences between which are sufficiently small (e.g. 2 members initialized every day of the month, ; this is the case for Met Office system)

Among all the systems that contribute to the C3S seasonal forecasts, some of them have opted to use a "burst" mode, while some others lag the start dates of the members of their ensembles. For more details, refer to the table below in the "Production schedules" subsection.

Fixed vs. on-the-fly hindcasts

Due to For several reasons, from computer load balance to flexibility in the introduction of changes in the systems, the different seasonal forecast contributors to C3S use different schedules to produce their hindcast sets:

• Fixed fixed hindcasts: . Some systems are designed so their expected lifetime will be around 4-5 years. Thus, once Once the system has been designed and tested, its development gets frozen and exactly that version of the model is used to run all the hindcast members ensemble hindcasts for the whole climate period for that model. In that way, the hindcasts are produced reference period are run. The advantage is that this reference dataset is available well in advance the of real-time forecasts and they constitute a fixed dataset during the lifetime of that seasonal forecasting systembeing issued, and its properties (biases, skill) can be quantified once for repeated use. As this is a very expensive exercise, it cannot be repeated too often and thus the system remains fixed for a long period of time.
• onOn-the-fly hindcasts: . Some other systems run the necessary set of hindcasts every time they produce systems prioritise more frequent upgrades, which means that the hindcast sets have to be run more frequently. To achieve this in practice, the full hindcast set is run every time a new real-time forecast . They are produce just is produced, slightly in advance (a few weeks) of the real-time forecast and using exactly the same version of the forecasting system. In this way changes in the system can be introduced more frequently and the computing of the hindcasts can be spanned over a longer period, then balancing the load of the computing resources.

Production schedules in the seasonal forecasting systems contributing to C3S

In the following table, it is shown the information about the ensemble sizes, start dates and production schedule for the seasonal forecasting systems contributing to C3S.

...

51 members

   26 start around the 22th
   25 start around the 15th

...

15 members

start dates?

...

51 members

   1 starts on the 1st
   24 start on the 25th
   24 start on the 20th

...

25 members

   1 starts on the 1st
   12 start on the 25th
   12 start on the 20th

...

7 members on the 1st
7 members on the 9th
7 members on the 17th
7 members on the 25th

...

on-the-fly

produced around 4-6 weeks in advanced

...

• This also offers the advantage of balancing the requirement for computing resources, but the compromise is the regular change of the model climatology.

^(a) Despite they are produced in a lagged mode, the data from MeteoFrance system5 is produced and provided at CDS as if all the members were initialized on the 1st

^(b) The production schedule of the MetOffice forecasting system doesn't prescribe how to build an ensemble for an specific nominal start date. The following choices are currently in use for the data archived in C3S:
FORECASTS: 50 members starting on or before the 1st of the month (NOTE: The original daily/subdaily data from all the daily members, and not just those 50, is processed and made available at C3S)
HINDCASTS: 28 members (7 starting on the 1st of the month and 7 on each of the 9th,17th and 25th of the previous month)

^(c) Due to the flexibility of MetOffice forecasting system, incidences on a given data are not usually recovered re-running the missed forecast but incrementing the number of members in one of the following days.
Example: An incidence happened affecting the 22nd of August/2017 forecast so no members are available for that date. Instead, there are 4 members available starting on the 23rd of August

^(d) CMCC seasonal forecast system routinely produces 80 real-time forecast members but just 50 of them are made publicly available as their contribution to C3S