Description of the model climate used for the Extreme Forecast Index

• General Introduction
• EPS Control model climate
• Unified VarEPS-monthly model climate
• Further Reading

General Introduction

For the production of climate-based forecast products it is essential to use a reliable model climate that reflects the long term characteristics of the underlying forecast model as closely as possible. For example, a different representation of the orography or the land-sea mask in the climate could directly introduce a significant EFI bias over mountainous or coastal areas. Aside from the requirement for a consistent orography and resolution, it is also desirable to have the same physical parameterisations, in order to introduce similar model error characteristics and therefore guarantee that any difference in the EFI values (or any other climate referenced product) is indeed the result of dissimilar forecast situations.

The WMO defines a climate as a collection of statistics over three decades (30-year period). It is impossible to fulfil this requirement by using only operational forecasts, because of the serious inconsistency between forecasts from different periods due to countless model upgrades. To use model reanalyses, such as ERA-40, is not an ideal solution either, because of the inconsistency between the model cycles being used to prepare the reanalyses and the actual operational forecasts. The real solution for the problem is therefore to rerun always the operationally used model (for as many years as possible), based on reanalyses as initial conditions. However, this places very high demands on computational resources, particularly for the EPS, where numerous forecasts need to be run each day.

A very important characteristic of a model climate is the sample size. To estimate the mean characteristics (such as the median) of the underlying climate distribution, we can rely on relatively small data sets. Considering a day specific model climate, in order to have a reliable mean, even the 30-value climate, having one value from each year for the specific day is sufficient. However, using a small sample to determine the tails of the distribution, where the rare events lie, would result in large sampling noise. To adequately sample the tails, we need significantly more realisations of the model. One way of increasing the sample size is to sample climate elements from a certain period (e.g. 1 or 2 weeks before and after) rather than just taking the single day for which we prepare the climate.

EPS Control model climate

The development of a medium-range model climate, used especially for the EFI, started in 2004, and it was implemented operationally on 1 February 2006 (with the latest resolution upgrade of the main model). It is based on an EPS Control re-forecast system. Every day at 12 UTC (instead of a “once per every model upgrade” the climate elements are prepared daily) 30 short range model integrations are made up to T+48 hours for the same day in each year from 1971 to 2000. The forecasts start from the interpolated ERA-40 analyses. Forecasts to 48-hours are the minimum requirement to provide a model estimate of the weather conditions for each day (for precipitation a 24-hour period 06-06 is needed).

Although the re-forecast system provides the climate for each day, the number of available forecasts for a single day is only 30 (years 1971 to 2000). To have a large enough sample, the EFI climate is composed with a 31-day sampling window using re-forecasts from 15 days before and after this day. The underlying assumption is that the daily climates of days not more than 15 days apart are reasonably similar and the re-forecast elements can be considered as realisations from the same climate distribution. With this choice the climate distributions are determined based on 930 (30 years × 31 days) re-forecast fields. This sample size seems to be large enough to represent extremes, such as 1st or 99th percentiles, reasonably smoothly even for precipitation or 10m wind which have very large intrinsic variability.

Beside this version of the EFI related model climatology, the Monthly Forecast System has always used a specific model climate. A set of model re-forecasts is prepared alongside the weekly forecast runs. 5 EPS members are run for the corresponding Thursday for the latest 12-year period. This 60-member climate mainly serves as a reference in order to remove model biases, drifts developing during the longer model integration.

Unified VarEPS-monthly model climate

As part of the unification of the VarEPS and Monthly forecast systems, a new and generic model climatology has been developed in order to fulfil the requirements for both the Monthly System and the medium range EPS including the EFI, keeping in mind the constraints imposed by the available computer resources.

The new model climate was developed on the basis of the existing monthly climate system. Unified VarEPS re-forecasts run once a week, from 00 UTC Thursday, when the monthly extension (Leg 3) is run. 4 EPS members plus the Control run for the full monthly forecast range up to 32 days, initialised from same day and month for the most recent 18 years. Therefore altogether 90 re-forecast files are available from one weekly run (18 years × 5 EPS members).

The new model climate for the EFI is based on 5 consecutive weekly re-forecast data sets (5 Thursdays), where the middle Thursday is the closest preceding Thursday to the actual EPS operational run date. This 4-week period is similar to the 31-day sampling window applied in the current system. Tests have confirmed that, especially in transition seasons, a model climate based on an asymmetrical sampling window (e.g. only 5 preceding Thursdays) can introduce large biases in the EFI. Therefore the re-forecasts need to be prepared 2 weeks in advance. The model climate files will be updated only once a week, after the new re-forecast fields (for a Thursday 2 weeks ahead) are available. For any model climate dependent application which is prepared on a daily basis, such as the EFI, the climate prepared for the latest Thursday can be taken. Although this configuration is not a properly symmetrical one, the maximum of 6-day shift between climate (always a Thursday) and actual EPS run dates introduces only small biases.

The old (currently operationally used) EFI climate changes day by day and the re-forecasts actually reflect the daily weather characteristics seen by the model with the closest possible accuracy and also has a reasonably large sample size of about 1000 values. On the other hand, it is the same for all forecast ranges (only short integration) and it is made only of the EPS Control therefore it is not capable of reproducing EPS characteristics. The disadvantage of the new system is that it samples the weather only on Thursdays, and even if it consists of a 5-member EPS, the potential climate sample sizes are noticeably smaller then 1000. However, there are two main advantages of the new system over the old one. It can provide forecast range specific climate and by having 5 EPS members it can simulate EPS characteristics such as spread and mean.

Further Reading

• Recent developments in extreme weather forecasting
  ECMWF Newsletter No. 107 (pp8-17) - Spring 2006