Note: HRES and Ensemble Control Forecast (ex-HRES) are scientifically, structurally and computationally identical.  With effect from Cy49r1, Ensemble Control Forecast (ex-HRES) output is equivalent to HRES output where shown in the diagrams.   At the time of the diagrams, HRES had resolution of 9km and ensemble members had a resolution of 18km.

2m Temperature errors

Bias

In general, temperatures are forecast fairly well over the globe.  On average, systematic errors in forecast 2m temperatures are generally <0.5°C.

Biases in 2m temperature (verified over land) vary with:

  • geography.
  • altitude.
  • season.
  • time of day.

Larger biases and errors occur in:


Diurnal Range of temperatures

The amplitude of the diurnal cycle is generally underestimated over land (a deficiency shared by most forecasting models).  This is especially the case in Europe during summer when the underestimation of temperature range reaches ~2°C across large areas.  Near-surface temperatures are generally too warm during night-time and slightly too cold during the day.  However, the degree to which the amplitude of the diurnal cycle is underestimated depends on region and season.  Night-time 2m temperatures are about 1–2°C too warm and surface temperatures about 2°C too warm.

Diurnal temperature changes are strongly influenced by incoming and outgoing heat flux. This is principally governed by the extent and thickness of cloud cover.  Model analysis and forecast of cloud and fog can have a strong impact on forecast errors.  

Most large errors seem to occur when the surface temperature is very cold and the lowest levels of the atmosphere may become extremely stable.  Tiny amounts of energy can correspond to large temperature changes at the surface because there is no convection to mix energy through the lower atmosphere.

Temperature errors often don’t depend strongly on the forecast range.

  

Effects contributing to temperature errors

Near-surface temperatures are related to a variety of processes:

  • cloud cover, cloud optical properties and aerosol 
  • albedo and radiative transfer
  • precipitation
  • surface fluxes
  • turbulent diffusion in the atmosphere
  • strength of land-atmosphere coupling
  • soil moisture and temperature

Some of the above processes in turn depend on land surface characteristics (vegetation, soil type, soil texture, etc.) and processes.

Errors in development and representation of near-surface inversion and/or low cloud cover is influential regarding errors in forecast surface and 2m temperatures.  These conditions are more likely with high pressure and cold, calm conditions and errors are relatively commonplace in such circumstances.  

It is vital to compare the observed and forecast thickness, the extent of low cloud, and the temperature and humidity structure of the lowest atmosphere.        

 

Coastal effects

The ground surface temperature (skin temperature) and surface albedo over land are very different to those over the water.  The land-sea mask defines whether the grid points are land or sea points.  But in coastal areas grid points will not capture the detail of the coastline.  Hence, surface radiative fluxes computed over the ocean may also be used by the IFS over the adjacent land.   This is because, for reasons of computational cost, the radiation code has to be run on a grid that is 6 times more coarse than the operational model grid.  This can lead to large near-surface temperature forecast errors at coastal land points.  To combat this problem the radiation code was changed and involved modifying the surface albedo when radiation calls are made. This leads to more to realistic coastal land temperatures.  Discussion of the land-sea mask and meteograms relates.

Despite the above IFS improvements, coastal temperatures still need to be viewed with caution, especially where urban areas are next to the sea.  The local drift of surface air from a land or sea source may differ significantly from model forecast low-level winds.  This can be especially true where the orography is complex and influence the actual, analysed, and forecast low-level winds.  Users should consider these aspects when assessing coastal temperature forecasts (e.g. forecast Southern European coastal temperatures have been observed, at times, to be too low during a Mediterranean heatwave).

Urban effects

Extensive concrete and buildings can possibly provide a source of heat (the heat island effect) and even moisture (from air-conditioning units).   Towns and cities are likely to have very different characteristics from other HTESSEL tiles which describe natural land coverage.   An urban tile in HTESSEL (introduced in Cy49r1) models the fluxes of heat, moisture and momentum and their effects around towns and cities.  Forecast screen temperatures in large urban areas, particularly cities and especially coastal cities, can still be a little low when compared to observations.   In particular, forecast screen temperatures can be too low on relatively clear, calm nights, and in winter where the urban area is surrounded by snow cover.   Users should assess the potential for deficiencies in low-level parameters and adjust as necessary.

Cloud effects

Effects of cloud cover

Analysed or forecast cloud cover has a large impact on forecasts of 2m temperature causes.  Cloud cover can hinder or enable radiative cooling (or also heating by insolation) during the forecast process.

Commonly:

  • too little cloud cover encourages:
    • night-time radiative cooling.  This results in significantly lower forecast 2m temperatures, especially over snow.
    • day-time radiative heating.  This results in higher forecast 2m temperatures.
  • too much cloud cover discourages:
    • night-time radiative cooling.  This results in anomalously high forecast minimum 2m temperatures.
    • day-time radiative heating.  This results in lower forecast 2m temperatures.

Much of the cold bias of night-time 2m temperature south of 60°N is associated with an underestimation of (low) cloudiness.  Wintertime night-time bias in Central Europe is smaller for occasions where (nearly) cloud-free conditions have been forecast and observed.

Day length can also be important.  At higher latitudes, cooling during the long nights may not be offset by solar radiation during the short days leading to a gradual day-by-day lowering in 2m temperatures.  

It is also possible, though less common, to have too little cloud in the forecast yet with temperatures that are too high!  These more unusual winter-time error scenarios commonly build up over a period of time.

Other effects of cloud

Analysed or forecast other cloud parameters can also have an impact.  Errors in the prediction of the temperature structure have a strong influence on forecast cloud layer(s) and on humidity forecasts, particularly in the lowest layers (Fig9.2.1.1-1 and Fig9.2.1.1-2).

Forecasts are influenced by incorrect:

  • optical depth. 
  • cloud type.
  • base height.
Verification of cloud

In order to assess how well the cloud has been captured forecasters should compare observed and forecast:

  • cloud cover (from observations or satellite pictures) with cloud analyses or forecasts.
  • cloud structure (from observed and background vertical profiles).


Fig9.2.1.1-1: The example illustrates how small errors in the extent and thickness of model forecast low cloud and fog can have large impact on forecast 2m temperatures.   Over Spain, 12UTC temperatures approached 12°C in areas where low cloud broke and cleared but remained cold where low cloud and fog persisted (subfreezing in some areas of freezing fog).  The model forecast low cloud to be variable in thickness and rather greater in extent than in reality.

Thus over Spain at 12UTC there were:

  • relatively large positive temperature errors (model forecast too warm, yellow areas) where fog and low clouds broke and cleared in the model but persisted in reality.  Too much cloud implies less insolation and persisting cold temperatures.
  • relatively large negative temperature errors (model forecast too cold, blue areas) where fog and low clouds broke and cleared in reality but persisted in the model.  Breaks in cloud cover allows temperatures to rise.

The forecast vertical profiles show that the model was fairly good at representing temperature inversions.  However, the model locally missed clouds and fog by a small margins.  Such small errors had a large impact on surface temperatures.  The warm area in central northern Spain associated with high ground above the overnight inversion.


Settled anticyclonic weather is often dry but cold, with persisting low clouds and/or fog in some places and sunny and warmer conditions in others.  Vertical profiles often show IFS struggling to represent low level anticyclonic inversions.  The model tends to capture low level cloud cover reasonably well but vertical profiles suggest that cloud layer is shallower in the forecast than actually.  This can mean the model cloud breaks too easily allowing surface and 2m temperatures to rise too much.  In these conditions temperatures can be some 2°C to 5°C too warm under clearing skies, but can be as much as 10°C where low cloud or fog is particularly reluctant to break or lift.

Generally, the shorter the forecast period the better temperature forecasts are, but even at T+12h errors can remain widespread and still large in few places.

  

Fig9.2.1.1-2: Charts showing distribution of low cloud over Europe at 12UTC 26 Dec 2024.  Also shown are actual and forecast (T+12) vertical profiles for Essen and Muenchen and the errors between the forecast and observed 2m temperatures more generally.  The largest positive errors (reds and yellows, model T2m too high) are where the model has cleared the cloud too quickly and allowed incorrect insolation.   The negative  errors ((blues, model T2m too low) are where the model has maintained the cloud too and hindered insolation.  The cloud distribution is critical and users should anticipate likely deficiencies by monitoring available surface or satellite information. 

Snow cover effects

The analysis and forecast of snow depth, snow compaction, and snow cover are important for forecasting lower-layer and near-surface temperatures and they can have a significant impact on forecast accuracy.  However, the relationship between snow cover and temperature is complex:

  • Where too much snow is analysed or forecast then forecast 2m temperatures may drop too low.
  • Where insufficient snow is analysed (or cleared too readily by the forecasted passage of weather systems) then forecast 2m temperatures may be too high.  
  • In marginal snow situations, when precipitation at the surface comprises both rain and snow, the snow component accumulates as lying snow whereas in reality it would usually melt instantly.  This can lead to the proliferation of areas where a there is an incorrect small amount of snow cover.
  • Even if snow-cover is reasonably well-represented large errors in forecast 2m temperatures can still occur.

At a given location, at a given time, all the snow on the ground is assumed to have the same density.  This density varies according to:

  • its age (the model facilitates slow, natural compression),
  • melting (small amounts of snow on the ground tend to take too long to melt, even if the temperature of the overlying air is well above 0°C).
  • interception (of rain)
  • addition of new snow.

The albedo is related to the extent (and age) of snow cover and snow characteristics in analysed and forecast fields have an effect on the radiation that could be absorbed.  This has a corresponding impact on forecasts of 2m and surface temperatures.    Better assessment of the albedo when the multi-layer snow scheme is introduced will allow faster response to changes in the radiative forcing.

Forecast temperatures can be in error by:

  • as much as 10°C too warm (very occasionally even worse) where there is:
    • snow cover under anticyclonic conditions and light winds,
    • strong radiation under clear skies leading to a strengthening night-time surface inversion.

(Both the above scenarios are difficult to forecast, but Finland, Scandinavia and large areas of northern Europe are particularly prone to forecast temperatures being too warm).

  • as much as 5°C or 10°C too low where there is:
    • spurious snow cover (i.e. snow that should have melted).  

(This effect is particularly evident where the model retains a snow depth >10cm, the level at which the ground is assumed to be completely covered).

Currently snow is modelled by a multi-layer snow scheme allowing a fairly realistic heat transfer.

(Note: Previous to June 2023 (Cy47r3 and earlier), only a single layer snow model was available.  There was no mechanism to deal with density variations in the vertical within the snowpack.  This had an impact on energy fluxes which in turn had potential to adversely affect the forecasts of 2m temperature.   For example, when new low density snow falls onto old dense snow, the atmosphere might be "re-insulated" from a ground heat source, allowing 2m temperatures to drop lower in reality than in the model.  In practice this particular problem will be exaggerated by temperature sensors ending up closer to the snow surface when snow has fallen (assuming they are not elevated manually)).


Fig9.2.1.1-3: The snow depth in the vicinity of Murmansk is shown as a shade of green (5-10cm).   A snow depth of 10cm (actual snow depth, not water equivalent ) is the threshold for the IFS to assume the entire grid box fully snow covered (snow cover fraction = 1 ).  Thus a difference around this threshold value can change the tile partitioning and thus snow coverage may not be uniform or continuous over the grid box.  The snow-free tiles would have less insulation from the soil underneath so maintaining the average skin temperature to higher temperature compared to a fully snow-covered grid box.  This can potentially impact the 2-metre temperature computation.

In this example a single layer snow model was used in Cy47r3 and before.  There is a significant difference between the observed (black) and forecast T+72 Ensemble Control Forecast (ex-HRES) forecast temperature (red) structure at Murmansk (location shown by the arrow).  The observed structure is much colder than that forecast, and in this case, surface snow cover appears to have been critical to the forecast.  The observed temperature structure could be due to stronger radiative cooling due to more extensive and/or deeper snow cover than is indicated in the IFS snow depth chart.  A multi-level snow model is used in Cy48r1 and later.

Also, notably, at 12UTC the observed structure (black line) shows more cloud in reality than forecast and yet is still a lot colder. 


Turbulent Mixing effects

Biases in near-surface temperatures during winter conditions are very sensitive to the representation of turbulent mixing in stable boundary layers.  The partition of momentum transport between dry and moist updrafts is particularly important.  Comparison with radiosondes in the lower 200m of the atmosphere suggests underestimation of the temperature gradient.  This is particularly pronounced at lower latitudes.  Full resolution of the details of the temperature structure in the lowest layers of the atmosphere is not possible with current computational resources.

Too much mixing increases the upward diffusion of heat,  hence reducing stability and/or temperature inversion and consequently the temperature fall at 2m and at the surface.  Errors tend to be much larger during low level inversion situations.  Note:

    • low level inversions are particularly common at high latitudes in winter.
    • the closer the inversion is to the surface, the larger is the potential error (forecast surface and 2m temperatures too high).
    • errors in wind profiles and/or wind directions relate to mis-representation mixing in convective boundary layers.

Vegetation effects

Temperature  errors (particularly in biases during spring and autumn) are in part related to the representation of vegetation (in terms of cover and seasonality), and evaporation over bare soil.  Heat flux from bare soil is also problematic.   Soil temperature and soil moisture is modelled in IFS but there is not a great deal of directly measured observations available.  However, the impact of heat and moisture fluxes can be a significant contributor to 2m and surface temperature errors, and hence have an impact on humidity.

Leaf area index is a measure of vegetation coverage and determines the degree of shading and how much radiation is absorbed or reflected.  Leaf area index varies in the model, month by month.  However, the leaf area index will not be representative if there is anomalous weather e.g. wind storms may strip leaves from trees, widespread fires may clear vegetation (and change the albedo).

Orography effects

IFS model orography smooths out valleys and mountain peaks, especially at lower resolutions.  A forecast 2m temperature may be unrepresentative if it has been calculated for an altitude significantly different from the true one.   A more representative height might be found at one of the nearby grid points.  Any remaining discrepancy can be overcome using Model Output Statistics (MOS) or statistical post-processing (see additional sources of information below).

The aspect of a location (i.e. orientation relative to the sun) is not taken into account.  Thus:

  • more solar energy will be gained on south-facing (N Hem) slopes.  This implies actual temperatures may be higher than forecast.
  • less solar energy will be received on north-facing (N Hem) slopes.  This implies actual temperatures may be lower than forecast, particularly where they are in shadow for much of the time and in sheltered valleys.

Overnight 2m temperatures tend to be too cold over rugged or mountainous areas.

Fig9.2.1.1-4: Example of temperature errors in mountainous areas. Forecast temperature at high altitude stations can be far too low due to mis-representation of the temperature inversion and over-development of night-time surface inversion.   Austria, DT 12UTC 27 Jan 2024, VT T+60 00UTC 30 Jan 2024.              

Large negative errors in night-time 2m temperatures (the forecast is too cold) can be present over the Alps or other mountainous areas.  A ridge/anticyclone can be associated with weak winds, large scale subsidence and a warm air mass.  Where such a ridge/anticyclone dominates over a mountainous area there can be large variations in temperatures at adjacent stations at very different altitudes.  Deep surface inversions of 2-metre temperature can develop. Therefore temperatures in valleys drop considerably below freezing whilst they stay much higher even positive on mountain tops.  Temperatures high up can be positive day and night whilst in the valley inversions formed with a well-marked diurnal cycle – sub-freezing temperature at night in some places as low as -7°C / -8°C but warmer at day time.  And the largest model errors tend to occur in the mountain tops where the model cannot represent well as the model has a smoother orography than reality.  As a result, especially with snow cover, the model builds sharp temperature inversions, even high up in the mountain where the air mass is warm.  The model is prone to very large temperature errors – errors are large at mountains tops as the model builds inversions which actually do not exist in the real world. 


 Fig9.2.1.1-5: An example of 2m temperatures in wintertime anticyclonic conditions in complex terrain.  Temperature analysis at 09UTC 30 Dec 2024 covering southern Germany, Austria, eastern Alps and the Tirol (from INCA analysis).  Temperatures are around -10°C in valleys and at the same time up to +10°C on high ground above the low-level cold pools. The spectral representation of model orography means detailed features such as these simply cannot be represented in a model with 9km resolution leading to large forecast errors in temperatures at some stations.   Note also the fanning out of some of the coldest air as it exits the Inn valley southeast of Munich.  (INCA Central Europe is the Integrated nowcasting system for the Central European area).

Lake effects

The effect of lakes is parameterised using FLake and a lake cover mask.  The sub-grid detail may not be completely captured and the energy fluxes may well be incorrectly estimated, particularly where frozen lakes are plentiful and/or forecast snow cover is uncertain.   These aspects can:

  • amplify errors in forecast 2m temperatures, and
  • introduce biases...

There are a number of complications which are not fully understood regarding the influence of frozen or snow-covered lakes upon the forecasting of low-level and 2m temperatures.  Lake temperatures can have a significant effect on forecast of temperatures, particularly where there is uncertainty whether the lake is frozen or not.  The impact can be significant leading to some uncertainty in forecast 2m temperatures in areas with many lakes.   Some high latitude areas where lakes are plentiful and prone to winter-time freezing are:

  • NE Scandinavia: mainly Finland, north Sweden.
  • Russia: Mainly West of the River Yenisey, also River Lena valley, parts of NE Siberia.
  • Canada: Mainly east of the Rockies and particularly: Labrador, Quebec, Ontario, Manitoba, Saskatchewan, Nunavut, Northwest Territories.
  • Alaska: Low lying areas.
  • Possibly some low lying parts of Southern Argentina.

Low Level Winds and Precipitation effects

If winds are light, melting of falling snow and/or evaporation of falling rain or snow, can cause local cooling down to surface levels.  Significant 2m temperature errors may develop if aspects of precipitation are not well captured by the model.  

In areas that are not completely flat, any stronger winds will tend to combat this tendency.  Low-lying areas can be warmer than would have otherwise been in the case (e.g. by re-establishment of a vertical lapse rate via adiabatic warming or cooling during descent or ascent over topography).  However, there is the further risk of turbulence cloud developing which can also decrease radiative cooling in the lower layers.

Persistent or heavy rainfall can produce waterlogged soil or flooded areas which will retard temperature rise.


Effects of aerosol and dust

Aerosol advected across a region can reduce incoming radiation.  Aerosol Optical Depth (AOD) measures the extinction of a ray of light as it passes through the atmosphere.  This can be due to advection of dust etc.  A very crude rule of thumb is that an anomaly (with respect to climatology) of 1 AOD unit corresponds to a 0.5-1.5 °C day-time temperature decrease under otherwise clear skies.  Cloud cover has a much stronger effect upon surface temperature   and mask any signal from the aerosols.  The radiative impact of the forecast aerosol value is more distinct for shorter lead-times (12 or 24 hours).  At longer lead times, the evolving differences in flow patterns and clouds may become more important for the surface temperature differences than the reduced solar radiation.  More information on aerosols and greenhouse gases is given elsewhere in the guide. 

 

Fig9.2.1.1-6: Example of forecast error associated with passage of a zone of associated with saharan dust.

Miscellaneous

  • Forecast maximum 2m temperatures can be too low particularly during anomalously hot weather.
  • If the predicted humidity is too low then maximum temperatures can be forecast to be too high.
  • Post-processing (e.g. using a calibrated statistical technique) usually improves 2m temperature forecasts, sometimes substantially.
  • Model 2m temperature output corresponds to short grass cover, because by meteorological convention observations are ordinarily made over such a surface.  This strategy may not work so well in:
    • complex terrain - e.g. forests with clearings.
    • over snow areas.  The algorithm which derives 2m temperature uses the model surface temperature while the snow surface is above the earth's surface.  This can be an important consideration where there is deep snow.  See section on snow effects.  

Suggested considerations to offset temperature errors

The forecaster should assess the potential for error due to the above factors by:

  • comparing analyses of temperature, dew point and soil moisture with observed data
  • assessing future "background" conditions and the potential impacts thereof (e.g. snowfall or cloud cover that might be different from atmospheric model predictions).


When assessing a forecast of 2m temperatures, users should try to assess whether:

  • the snow-cover is adequately analysed or forecast, especially taking into account:
    • any observations of snow cover and density,
    • the tendency for insufficient melting of snow on the ground.
  • the cloud cover is adequately analysed or forecast.
  • the boundary layer temperature (and humidity) structure is adequately analysed.
  • wind strength is adequately modelled.
  • the temperature and moisture structures of the lower atmosphere are well represented.

Thus if:

  • too little snow-cover and/or too much cloud is analysed then there is a risk forecast temperatures may be too high.
  • too extensive snow-cover and/or too little cloud is analysed then there is a risk forecast temperatures may be too low. But in the case of too little cloud or more wind sometimes temperatures may be too high.
  • the boundary layer structure is not successfully analysed then there is a risk forecast temperatures may correspondingly be in error.
  • winds are too strong or too weak then forecast temperatures may have larger errors (particularly at high latitudes in winter where the role of insolation in offsetting radiative cooling is minima.

(FUG Associated with Cy49r1)


  • No labels