These statistics are calculated over all model runs. Stratifying the statistics by time of
day reveals additional interesting information about their respective performance
characteristics. In the example of Fig 11.30, the wind error statistics are computed for all
model runs initialized at 00 UTC, for the 01 UTC runs, etc., and the resulting time series
of 12-h error statistics graphed. Note that MM5 and WRF forecasts initialized between
06 and 11 UTC have lower bias errors than model runs initialized at other times. These
forecasts are valid between 18 and 23 UTC, which is 11 AM to 4 PM CST. During these
hours, daytime heating mixes and homogenizes the lowest part of the atmosphere and
increases the effective friction encountered by the air. This process links upper-level
flow with surface wind and enhances predictability. At night, the near-surface air is
effectively decoupled from the flow aloft by radiative cooling. This allows the surface
air to more easily respond to small, subgrid-scale forcing effects such as radiation-driven
drainage flows, which the models do not accurately predict. The errors are caused by
overestimation of the frictional effect, resulting in positive wind speed bias at night.
133