Short-Term Explicit Prediction

Background

The Short-Term Explicit Prediction (STEP) Program is a multi-NCAR Laboratory activity with the overarching goal to improve the short-term (0-36 hours) forecasting of high-impact weather events such as severe thunderstorms (heavy rain, tornados, downburst, flash flood, lightning and hail), winter storms (snow, freezing rain and drizzle), and hurricanes. The STEP program emphasizes several research areas that are crucial for advancing the science and application of the short-term prediction of high-impact weather, through collaborative effort incorporating national and international scientists, engineers, and operational personnel from universities, government institutions and the private sector. Most of the forecasting/nowcasting systems and analysis tools developed under STEP are available to the communities for the support of research and real-time operations.

Figure 1. Flowchart for the STEP Hydromet Prediction System.
Figure 1. Flowchart for the STEP Hydromet Prediction System.

In FY19, RAL’s STEP effort emphasized development of nowcasting techniques based on advanced data assimilation and machine learning techniques, improving microphysics scheme for more skillful mail prediction, further improvement of WRF-hydro system. All three research areas are crucial components for the integrated Hydromet Prediction System (Figure 1). The overarching objective of these research efforts are to advance the prediction of heavy rainfall, flash floods and streamflow through the integration of state-of-the-art rainfall estimation, precipitation forecasting/nowcasting, and hydrology modeling techniques into one seamless system.

Error characterization of model precipitation forecasting

In the past year, efforts to improve data assimilation and nowcasting systems have been focused on evaluation of data from the recent STEP hydromet test bed experiment and from the PECAN field campaign. Some additional work supported the RELAMPAGO field campaign. 

Figure 2. Illustration of the parallel motion of a convective object identified in both observations (MRMS: orange) and WRF forecasts (green).
Figure 2. Illustration of the parallel motion of a convective object identified in both observations (MRMS: orange) and WRF forecasts (green).

RAL has evaluated the large collection of WRF 0-24 hour forecasts of convective precipitation in comparison to the EOL radar QPE product, and the NOAA MRMS precipitation product.  In these evaluations, the Method for Object based Diagnostic Evaluation (MODE) and MODE-Time Domain (MODE-TD) tools have been used to characterize more than just the error in precipitation fields.  These tools have been used to identify the spatial structure of the storms and evaluate weaknesses in the prediction of storm size, magnitude, location, and translation. In addition, raw grid-point statistical characteristics (e.g. mean square error and correlation) have been computed over the larger great plains domain.  In this region, we identified regions with weaker correlation between model and observed precipitation closer to the Rocky Mountain front range, and maximums in predictability (correlations up to 0.9) occurring in western Kansas.  Other patterns identified include a known tendency of WRF to over predict the number of light precipitation events, and under predict the number of heavy (>15mm/hr) events. The MODE object based verification analysis showed that errors in position, translation and intensity, had persistence for up to 6 hours (Figure 2).  This implies corrections to the position and translation of a convective system can be derived from comparisons to radar observations and applied to future WRF forecast period.  

Working with the streamflow forecasting group, several major case studies were identified for further investigation.  These cases include two major floods along Cherry Creek near Parker, one of which (June 13, 2016) resulted in streamflow two orders of magnitude greater than the background flow (Figure 3). Comparisons to streamflow are challenging because river systems are commonly modified by human infrastructure, and finding a basin with minimal human modifications, an operational stream, and major flooding events in the STEP hydromet test period was difficult.  The WRF forecasts for this period were shown to predict the evolution of a thunderstorm in the region, but failed to locate it with the correct intensity over the water shed.  The WRF-Hydro model was shown to be able to reproduce a major flood when reliable precipitation data were available, although some errors in timing were consistent regardless of the precipitation event, and work is ongoing to examine the tradeoff between hydrologic model parameters, precipitation errors to identify an ensemble of streamflow forecasts that can be used to make a more statistically reliable flood forecast.

Figure 3: Observed streamflow on Cherry Creek near Parker, CO for a flood event caused by intense convective rainfall on June 13, 2016, median flow for this period is 8 cubic feet per secton (CFS) and observed peak discharge was over 800 CFS.
Figure 3. Observed streamflow on Cherry Creek near Parker, CO for a flood event caused by intense convective rainfall on June 13, 2016, median flow for this period is 8 cubic feet per secton (CFS) and observed peak discharge was over 800 CFS.

In addition, RAL has completed the analysis of the refractivity fields derived from radar observations for the duration of the PECAN experiment and drafted a paper for publications.  This work has been presented, at the AMS radar conference in Japan in September 2019, and the paper will be an important contribution to the literature.

FY20 Plans

  • Develop machine learning algorithms to increase the information available to data assimilation systems.  This work will leverage the 3D patterns of radar reflectivity and background characteristics derived from large existing datasets of convection permitting WRF model simulations. 
  • Conduct OSSE forecast experiments to understand the possible forecast improvements if machine learning algorithms can be used to improve the initialization of wind fields around thunderstorms.
  • Develop and test algorithms to merge WRF forecasts of storm translation and intensification with radar extrapolation algorithms to improve 0-3hr forecasts. 
  • Examine GOES-16 data for improved detection and nowcasting of thunderstorms.  With the additional channels, higher spatial resolution, and higher frequency of imagery, GOES-16 has the potential for greater applicability to short-term nowcasting.   RAL will compare these observations to radar observations, EOL's micropulse DIAL, and the numerous soundings collected during the RELAMPAGO field campaign.   RAL will test machine learning algorithms to automatically identify robust, reliable patterns in the storm initiation and evolution using a combination of these observations. 
  • Further analyses of data from the RELAMPAGO field campaign to better understand convective initiation (CI) and severe weather occurring in Argentina along the Sierras de Cordobas. This will be compared and contrasted with the weather evolution along the Rockies and over the Great Plains.

 

Data assimilation to improve model-based nowcasting

Figure 4. Comparison of 2-h precipitation forecasts by FINECAST. (a) QPE; (b) forecasts from an experiment assimilating radar alone; (c) same as (b) but assimilating both radar and lightning data; (d) performance diagram for precipitation threshold of 14mm/2hr; and (e) ETS score comparison.
Figure 4. Comparison of 2-h precipitation forecasts by FINECAST. (a) QPE; (b) forecasts from an experiment assimilating radar alone; (c) same as (b) but assimilating both radar and lightning data; (d) performance diagram for precipitation threshold of 14mm/2hr; and (e) ETS score comparison.

In the past year efforts on data assimilation focused on the assimilation of new observations and development of a hybrid system that merges WRFDA with DART based on an ensemble variational data assimilation framework (EnVAR). A prototype  of the hybrid system is now ready for further testing. Recently we have started to apply the system to convective-scale radar data assimilation.

Studies on the assimilation of new observations include development of a technique for lightning data assimilation and evaluate the impact of rainfall (QPE) on convective-scale data assimilation and forecasting. Both types of new observations were simultaneously assimilated with radar observations and their added benefits were evaluated.

Figure 5. 6-hour accumulated rainfall forecasts from an experiment assimilating rainfall only (second from left), radar only (third), and rainfall and radar (right) for the Meiyu MCS case that occurred on June 2 2017.
Figure 5. 6-hour accumulated rainfall forecasts from an experiment assimilating rainfall only (second from left), radar only (third), and rainfall and radar (right) for the Meiyu MCS case that occurred on June 2 2017.

The lightning data assimilation method attempts to obtain updraft information from the observed lightning flash rate. Since the lightning flash rate is only correlated with the maximum updraft in an air column, a vertical velocity profile is needed to map the updraft in the whole column. We tested a few approaches to obtain the profile and evaluated their impact on the lightning data assimilation. Another challenge in lightning data assimilation is to understand the representativeness of the lightning observations. Since the lightning flash is an instantaneous and point quantity but NWP models have much coarser resolution spatially and temporally, it is important to understand the representative error of the model. We used a method to match the scale of the lightning data to that of the NWP model such that the lightning data can be effectively assimilated into the model. Our results show that the lightning data assimilation in combination with radar data significantly improves the convective-scale analysis and very-short-term forecasting. The new lightning data assimilation method have been implemented to both WRFDA 3DVar and the 4DVar-based nowcasting system FINECAST. Figure 4 shows the positive impact of the lightning data assimilation from FINECAST by comparing the combined radar and lightning data assimilation experiment (RAD+LTN) with  the radar alone experiment (RAD). The RAD+LTN experiment successfully predicts the newly initiated storm cell in front of a decaying convective system. The 2 hour forecast skill is significantly increased.

The assimilation of QPE data were conducted using WRFDA 4DVar. A case of MCS embedded within a Meiyu front that affected Taiwan with heavy rain and severe flood was used for the study. Our study shows that while radar observations are crucial to analyze the air motion and its associated convergence through data assimilation, it has not positive effect (sometimes negative effect) on the humidity field. We further demonstrated that the assimilation of QPE data can greatly improve the low- to mid-level humidity analysis, hence result in more skillful short-term heavy rain forecasts. Figure 5 shows that the addition of the rainfall data assimilation to radar improves the pattern, location, and intensity of the rainband over the Taiwan island.  

Figure 6. Water vapor analyses from WRFDA 4DVar on the eighth model level from the radar alone experiment (left) and radar + rainfall experiment (right).
Figure 6. Water vapor analyses from WRFDA 4DVar on the eighth model level from the radar alone experiment (left) and radar + rainfall experiment (right).

Figure 6 compares the humidity analysis fields between the radar alone experiment and the radar plus rainfall experiment. The impact of the added rainfall assimilation on the humidity analysis in the rainband region is significant, which is the main reason for the improved heavy rain forecast.

In the past year, we also further improved the performance of FINECAST including the improvement of surface data assimilation and terrain scheme, and adding a scheme to blend the advection only forecast and the full-model forecast to address the issue of model spin up. 

FY20 plans

 

  • Continue the testing and improvement of model-based nowcastingusing FINECAST
  • Design and test a hybrid convective-scale data assimilation system based on NCAR’s WRFDA and DART systems
  • Continue the studies to assess impacts of new high-resolution observations including cellphone pressure observations and radar refractivity observations (latter is in collaboration with EOL)
  • Collaborate with EOL on OSSE studies to evaluate observation system design concerning EOL’s new multi-pulse dial (MPD)

 

Evaluation and improvement of model microphysics parameterization    

Development of prototype multi-moment graupel/hail category in Thompson microphysics parameterization

Figure 7. Lowest model level simulated reflectivity at 1500 UTC on 20 June 2015 over South Dakota from a test of the new y-intercept diagnostic in the original (mp=28) Thompson single-moment graupel/hail scheme (left) versus that from a test of the new (mp=38) Thompson multi-moment graupel/hail scheme (right).  The maximum hail size is denoted by the contour lines; 0.5 cm in black, 1.0 cm in blue, and 2.0 cm in magenta.  The white lines denote the area for line-averaged cross sections shown in Figure 8.
Figure 7. Lowest model level simulated reflectivity at 1500 UTC on 20 June 2015 over South Dakota from a test of the new y-intercept diagnostic in the original (mp=28) Thompson single-moment graupel/hail scheme (left) versus that from a test of the new (mp=38) Thompson multi-moment graupel/hail scheme (right).  The maximum hail size is denoted by the contour lines; 0.5 cm in black, 1.0 cm in blue, and 2.0 cm in magenta.  The white lines denote the area for line-averaged cross sections shown in Figure 8.

In FY17-FY18, a multi-moment graupel/hail category was added in the Thompson microphysics scheme.  For this multi-moment graupel/hail category, number concentration and bulk volume mixing ratio are now predicted variables, in addition to the bulk mass mixing ratio. This allows for the graupel density to vary in space and time and be diagnosed from the new predicted variables.   Based upon initial testing of the new scheme, the code was refined and prepared for use in a real-time forecast model that was run during the RELAMPAGO field campaign in FY19.  In addition, hail pads were deployed during RELAMPAGO in order to obtain ground truth hail size measurements. 

Improvements to the single-moment graupel/hail category in Thompson microphysics

Figure 8. Line averaged cross sections of model simulated reflectivity from the two simulations shown in Figure 7.
Figure 8. Line averaged cross sections of model simulated reflectivity from the two simulations shown in Figure 7.

The original Thompson single-moment graupel/hail category only predicted graupel/hail mass mixing ratio, however it diagnosed the y-intercept of the size distribution allowing it to vary and not be prescribed as a constant value.  A recent study was published by Field et al. (2019) that showed observed relationships between the slope and y-intercept parameters of hail size distributions based upon in situ hail measurements from 18 flights of the T-28 armored aircraft between 1995-2003.  This study also revealed that the original Thompson single-moment graupel/hail diagnostic y-intercept values were far too large compared to observed values.  Therefore, using these data as guidance, new diagnostic relationships were developed and tested.  This work in ongoing and in parallel with the multi-moment graupel/hail scheme testing and evaluation, by using the same case studies and observations to compare with both simulations (Figs. 7-8). 

Established methods for baseline performance evaluation

In order to evaluate impacts on storm structure, QPF, and, in particular, storm evolution due to physics parameterizations, including updates made to the Thompson microphysics scheme, an object-based evaluation tool that tracks storms over time is needed.  The Method for Object-based Diagnostic Evaluation (MODE)-Time Domain (TD) is a tool developed at NCAR for such analyses, allowing users to set several parameters to evaluate storms with certain spatial scales and intensities.  Moreover, there are several parameters that users can set that impact how observed and modeled objects are matched for comparative evaluation metrics. In FY18-19, sensitivity tests were performed that varied the matching parameters to quantify the impact of these choices on the evaluation outcomes.  This effort has established methods for how to apply MODE-TD to quantify impacts of various physics choices, with specific emphasis on the new multi-moment graupel/hail category, and in order to determine baseline performance metrics.

FY20 plans

  • Evaluate the Thompson microphysics scheme updates to improve forecasted convective storm structure, evolution, and QPF.  This includes both the new y-intercept diagnostic in the single-moment graupel/hail scheme and the new multi-moment graupel/hail scheme.
  • Run a ~30-day simulation experiment to provide a longer-term evaluation of the schemes, as opposed to single case studies.  This experiment period will be selected to coincide with a complementary field experiment or numerical modeling exercise (i.e. RELAMPAGO and/or Hazardous Weather Testbed), to broaden its applicability and enhance the observations that are available for the study.
  • Use MODE-TD to evaluate the prototype microphysics scheme compared to the original Thompson scheme and other microphysics schemes and/or forecast ensemble members as available.
  • Use evaluation results to inform further microphysics parameterization improvements.