May 21 - July 30
Regime Dependent Verification and Calibration of a 10-Member Convection-Permitting Ensemble during the 2019 HWT SFE
Soleil Cotterell, Aaron Johnson, and Xuguang Wang
What is already known:
- Convection allowing ensembles (CAEs) can explicitly resolve convective scale features, thus the initiation and subsequent propagation of individual storms can be modelled.
- Computational approximations and parameterization scheme configurations introduce and influence forecast bias.
- CDF-based bias correction can improve the skill of neighborhood-based CAE forecasts.
- These biases can exhibit regional and atmospheric flow regime dependence.
- Correcting for these biases will lead to more skillful forecasts.
What this study adds:
- For May 2019 CAE forecasts, removing bias from meteorologically similar regions separately improved composite reflectivity forecast skill compared to region-blind bias correction.
- Self Organizing Maps (SOMs) effectively identify physically realistic synoptic patterns occurring during the May severe weather season over the last 10 years.
- Naïve use of SOM-derived synoptic classifications to remove bias from meteorologically similar synoptic flow regimes separately did not generally improve skill of the May 2019 CAE forecasts compared to regime-blind bias correction, though a significant increase in skill was observed within one regime.
- We hypothesize that further skill improvements are still possible using larger dataset sizes coupled with further optimization of the SOM classification methodology.
Abstract:
3km grid spaced forecasts generated during the 2019 NOAA Hazardous Weather Testbed Spring Forecast Experiment (HWT SFE) by the Multiscale Data Assimilation and Predictability group at the University Ok- lahoma are verified using the Neighborhood Maximum Ensemble Probability (NMEP). The verification was first performed on variable HWT defined domains and later extended to a large fixed CONUS domain. 24- hour forecasts of hourly maximum composite radar reflectivity initialized at 0000 UTC on 26 days during the spring of 2019 were evaluated. Forecasts on the HWT domains were generally skilled albeit with a no- table over-forecasting bias, while forecasts on the fixed domain were generally poor. Cumulative Distribution Function (CDF) bias correction was performed in each domain and forecasts were reverified. Further, the fixed domain was segmented into sub-domains and a novel regional CDF (RCDF) bias correction approach was undertaken. CDF corrected forecasts on the fixed domain were still poorer than climatology, but signifi- cantly more skilled than without calibration. RCDF corrected forecasts on the fixed domain were significantly more skilled than CDF forecasts and were the only forecasts to exceed climatological skill. Synoptic pattern classification using Self Organizing Maps (SOMs) identified physically realistic synoptic patterns occurring over a ten-year climatology. na ̈ıvely using the SOM-derived synoptic classification to remove bias from me- teorologically similar synoptic flow regimes separately did not generally improve forecast skill compared to regime-blind bias correction, though an interesting exception is noted. Suggestions are made for improving the robustness of the regime-dependent calibration scheme.