1. Improving forecasts through rapid updating of temperature trajectories and statistical post-processing
Nina Schuhen, Thordis L. Thorarinsdottir and Alex Lenkoski

2. Status quo Forecasts are not revisited once issued
Newest forecast run is usually the best
Idea: correct a forecast trajectory while it verifies

3. More skill!

4. Data set: UK Met Office MOGREPS-UK 2.2 km 70 Levels
36 hourly forecast 4 times/day
12 members
Interpolated to 149 observation sites in UK and Ireland
Training period: Jan to Dec 2014
Evaluation period: Jan 2015 to Jul 2016
Surface temperature

5. EMOS / NGR Raw ensemble output is calibrated using EMOS (NGR)
Removal of deterministic and probabilistic biases
Performed after the ensemble run is completed, based on a rolling training period
Gneiting, T., A.E. Raftery, A.H. Westveld, and T. Goldman, 2005: Calibrated Probabilistic Forecasting Using Ensemble Model Output Statistics and Minimum CRPS Estimation. Mon. Wea. Rev., 133, 1098–1118

6. Previous results for EMOS

7. EMOS / NGR Raw ensemble output is calibrated using EMOS (NGR)
Removal of deterministic and probabilistic biases
Performed after the ensemble run is completed, based on a rolling training period
16% reduction in CRPS
Here: correct EMOS mean, keep calibrated EMOS spread
Gneiting, T., A.E. Raftery, A.H. Westveld, and T. Goldman, 2005: Calibrated Probabilistic Forecasting Using Ensemble Model Output Statistics and Minimum CRPS Estimation. Mon. Wea. Rev., 133, 1098–1118

8. Correction of forecast trajectories
We allow for 1 hour delay for measurements to be transmitted, processed, quality-controlled, etc.
Here: 03 UTC run at Heathrow Airport
Data analysis shows no systematic trend in forecast error
Strong connection to diurnal cycle
Regression approach is the most appropriate

9. Framework
At any time a new forecast is issued for the next 36 hours: with
Observations become available while this forecast trajectory is valid
Compute forecast error at :
Correct the next forecast(s) using a function of this error:

10. Forecast error correlation matrix

11. Regression method Estimate future errors by linear regression
Previous forecasts’ errors as predictors:
Correct forecasts by adding the estimated error:
Find earliest lead time, where is significantly different from 0

12. Length of error correction period

13. Development of forecast error

14. Development of forecast error

15. Development of forecast error

16. Extension to multiple sites

17. Conclusions Strong correlation between error at different lead times
Predictability depends heavily on the time of day
Regression model reduces forecast error overall
Big margin for short-term corrections
Might prefer corrected forecasts from a previous model run to the first forecasts from the current one

18. Future work Investigate relationship between forecast runs
Develop rules for combining forecast runs to one «best» trajectory
Take into account seasonal effects
Kriging to spread regression parameters spatially?
Extend to other weather parameters

19. Example