Combining weather radar nowcasts and numerical weather prediction models to estimate short-term quantitative precipitation and uncertainty

The topic of this Ph.D. thesis is short term forecasting of precipitation for up to 6
hours called nowcasts. The focus is on improving the precision of deterministic
nowcasts, assimilation of radar extrapolation model (REM) data into Danish
Meteorological Institutes (DMI) HIRLAM numerical weather prediction (NWP)
model and produce quantitative estimations of nowcast uncertainty.

In real time control of urban drainage systems, nowcasting is used to increase the
margin for decision-making. The spatial extent of urban drainag
e catchments is very small in a meteorological context. This is a problem since smal
l scale features of the precipitation are the least predictable and hence very difficult to anticipate. This also leads to uncertainty at urban scale, which needs to be addressed.
Initially, Kalman filtering is used to stabilise the advection field in order to increase
the precision of a Co-TREC based REM. The filter is calibrated against atmospheric
observations of radial velocity measured by a Doppler radar. The results from
pooled skill scores from 16 events show only a slight improvement. The positive
contribution, from applying Kalman filtering, is increased stability computed by the
relative standard deviation.

A significant result of this Ph.D. study is major improvements in predictability of
DMI HIRLAM NWP model by assimilation of REM data. A new nudging assimilation method developed at DMI was used to assimilate the REM data. The assimilation technique enhances convection in case of under-prediction of precipitation and reduces convection in the opposite case. The result is based on evaluation of 8 events from august 2010 and an extreme event from 2 July 2011. Both spatial predictability and accumulated volumes benefit from the REM data assimilation. The system is currently being tested at DMI to become an operational system.

To address the uncertainty of REM nowcasting, a new ensemble prediction system
was developed called RESEMBLE (Rainfall Extrapolation System – EnseMBLE).
The novelty of this method is the separation of advection – and evolution
uncertainty and the way the temporal correlation is incorporated by a numerical
interpolation technique. The results demonstrate that ensemble mean performs with
higher correlation than the deterministic prediction compared to observations. The
system, with good skill, is able to predict the location and intensity of precipitation
and the ensemble spread is in proportion to the uncertainty of ensemble mean. The
system is also tested as input for an urban drainage system with promising results.
The encouraging results from assimilation of REM data into DMI HIRLAM NWP
model also inspired the work of initiating HIRLAM NWP ensemble members by assimilation of REM ensemble members. The same nudging assimilation technique
was applied to assimilate ensemble members from RESEMBLE into the NWP
model. The results showed a rapid initiation of ensemble members, reasonable
reproduction of nowcast uncertainty and a higher performance than ensemble mean
than runs without using RESEMBLE assimilation. A slight bias was also
demonstrated in the prediction towards high intensities but this was expected since
the model was tuned towards high intensity precipitation.