ACD ASR 2003 ASP ASR 2003 ATD ASR 2003 CGD ASR 2003 ESIG ASR 2003 HAO ASR 2003 RAP ASR 2003 MMM ASR 2003 SCD ASR 2003 Go to NCAR Go to UCAR Go to NSF Go to NCAR's ASR 2002


Convective Weather


C. Convective Weather Forecasting

[Background] [Field Deployments in FY 2003]
[Variational Doppler Radar Analysis System]
[Forecast Demonstrations]
[Quantitative Precipitation Nowcasting]

1. Background

Part of the challenge in forecasting convective weather is to be able to predict the initiation and location of new convection, i.e., thunderstorms, 30-min to 2-h in advance. Other parts of the challenge is to anticipate the merger of thunderstorm cells and predict the onset of thunderstorm dissipation. Over the past few years, effort has been focussed on developing automated thunderstorm nowcasting systems that combine information from observation based feature detections, Numerical Weather Prediction (NWP), and human forecaster input. Features or forecast parameters relevant to thunderstorm evolution include boundary-layer convergence zones, boundary-storm interactions, cumulus cloud detection and growth, boundary-relative shear profile, boundary-relative updraft strength and storm trends and tracks. The output from algorithms used to identify these features are combined using fuzzy logic in the form of user-defined membership functions, interest fields, and weighting schemes to produce forecasts of thunderstorm initiation, growth and decay. Two operational systems were deployed this year: a regional-scale forecast system called the AutoNowcaster (ANC) and a national scale system called the National Convective Weather Forecast (NCWF).

2. Field Deployments in 2003

The AutoNowcaster is now a permanent installation at the Weather Forecast Offices at the White Sands Missile Range (WSMR) in New Mexico as part of the Army Test and Evaluation Command (ATEC) 4DWX program.

A collaborative demonstration of convective weather forecast products for the FAA Aviation Weather Research program took place in summer 2003. NCAR ran forecast systems over Illinois and Indiana (implemented and supported by S. Dettling, N. Oien, T. Betancourt, R. Roberts, and T. Saxen), the Northeastern United States, and over the CONUS domains (N. Rehak, G. Cunning, D. Meganhardt, and C. Mueller). Each system varied in terms of input data, predictor fields, and resolution. Efforts to forecast initiation were limited to the Illinois/Indiana domain. All systems forecast storm location, growth and decay.

3. Variational Doppler Radar Analysis System (VDRAS)

The Variational Doppler Radar Analysis System (VDRAS) was modified and enhanced to assimilate data from multiple radars over a mesoscale domain (~500 km). A number of modifications to VDRAS were necessary to allow it to run efficiently on a larger, multi-radar, domain. The temporal and spatial smoothness coefficients, as well as the weighting given to the surface data compared to the radar data. A further enhancement to VDRAS by A. Caya further enhanced VDRAS by adding an evaporative cooling term to the dry model based on the observed reflectivity field. The inclusion of this term produced stronger cold pools in the region of convection, which in turn increased the low level convergence.

VDRAS was successfully implemented in the RCWF field program to perform low-level wind analysis with observations from four WSR-88D radars [Chicago (LOT), Northern Indiana (IWX), Indianapolis (IND) and Cincinnati (ILN)]. The analysis domain covered an area 600 km x 550 km with a resolution of 5 km. VDRAS ran robustly and produced smooth wind analyses throughout the field season. Wind shifts along convergence lines were displayed well in several severe weather cases.


Figure 3.1. An example of a VDRAS wind analysis on 26 August 2003. A large, isolated storm moved from west to east across the analysis domain. The low-level reflectivity field from the IWX radar is overlaid on the VDRAS wind analysis at 180 m, AGL (every second vector is displayed). The analysis captures the wind surge along the leading edge of the storm's gust front.

4. Forecast Demonstrations

a) Regional AutoNowcaster (ANC)

During the summer of 2003, D. Albo and T. Saxen developed and tested a new algorithm that allows a forecaster to enter boundaries (both moving and stationary) with minimal effort to be utilized in the forecast logic in real-time. Boundaries were entered using the mouse to draw a line on the display and proved to be a fairly straightforward and quick process. The system continues to extrapolate and display the extrapolated positions of the boundary until it is removed via a control window on the display. Boundaries tracked quite well and were seldom adjusted once entered. Most of the interaction was in the initial entering of the boundary and removal.

Forecaster-entered boundaries were usually frontal features that significantly affected convective development and evolution but were difficult for the automated algorithms to detect. These large-scale boundaries spanning ultiple radars had indistinct features associated with them (i.e., thin lines in the reflectivity field and/or sharp shear features). By using surface observations, satellite imagery, and VDRAS output in conjunction with the radar data, these boundaries were generally easy for the human user to identify.

Figure 4.1 shows time-series plots of performance statistics that illustrate the effects of the human inserted boundaries on the forecast. Boundaries were only entered during the hours shown on the plots. The boundaries had a dramatic effect on statistics, especially the POD during the late morning and early afternoon when a significant amount of growth and new initiation occurs. Increases in the POD of 10% to 40% were not uncommon during these time periods. It should be noted that the three different levels of initiation shown as blue, green, and black lines are initiation likelihood. The exact location of storm initiation is extremely difficult to forecast. The ANC algorithm forecasts regions where initiation is likely to occur. Thus, FAR for these initiation nowcasts is higher than expected from an extrapolation forecast. Roberts et al. (2003) prepared a paper for the 31st Radar Conference that illustrates a case where dramatically improved storm initiation nowcasts were obtained when the forecaster entered a synoptic front.

Figure 4.1. Time series plots of CSI (top), POD (middle) and FAR (lower) for four days: 18 June ; 29 June; 1 July; and 8 July. The black arrows in the figures indicate periods of forecaster-input boundaries. The lines represent persistence (yellow), extrapolation (cyan), and three levels of initiation likelihood (blue, green, and black) forecasts.

In the AutoNowcast system, the growth/decay and the initiation components of the forecast are handled separately. For the growth/decay part of the forecasts, the overall results from this summer were favorable. Failures and false alarms decreased while successes increased relative to the extrapolation forecasts. This was the case for 25% and 40% of the days (from June - August) for the north and south parts of the domain respectively. An example growth and decay forecast is shown in Figure 4.2. Growth (decay) areas are indicated where the white contours are outside (inside) of the green contours. The majority of the line of storms was forecast to grow, but some areas to the northeast were forecast to decay completely. Some convection on the back side of the convective line to the southwest was forecast to decay. The AutoNowcast forecast nearly closed the gap found near the center of the line. The decay trends at the northeast and the back side of the southwest parts of convective the line were quite accurate as were the growth areas in the other parts of the line.

Figure 4.2. The 2.5 km reflectivity CAPPI from 1557UTC on 29 August 2003 with extrapolation (green) and Auto-Nowcast growth and decay forecast (white) contours are shown on the left. The forecast performance image (forecast super-positioned on the reflectivity data at valid time) is shown on the right.

b) Probability forecasts over CONUS – NCWF

D. Megenhardt, N. Rehak, G. Cunning and C. Mueller ran an experimental version of the NCWF (NCWF-2) that provides 1- and 2-h probability forecasts for convection updated every 5 min (Fig. 4.3). Regions of growth are captured by use of Rapid Update Cycle (RUC) output along with radar trending and diurnal considerations. Dissipation is trended based on storm area coverage. During the summer, the NCWF was run in real time at NCAR using commercial vendor 2-km radar reflectivity data and national lightning detection network (cloud-to-ground) data.

NCWF-2 high probability regions are primarily the result of extrapolation and area trending. Several improvements have been implemented in motion vector calculations over the operational NCWF that runs at the Aviation Weather Center (AWC). The operation at NCWF provides motion vectors for storms > 512 km2 and have a history of at least 30 min. NCWF-2 uses RUC steering-flow winds to forecast motions for small storms when history is not available. This largely eliminates some of the criticisms of NCWF, including lack of forecasts when storms are too small, delay of forecasts when storms are growing and there is no storm history, and intermittent forecasts when the storm area straddles the threshold (512 km2). Another issue that plagues the operational NCWF is noisy motion vectors, especially during times of storm development and growth. NCWF-2 addresses this issue by inclusion of more rigorous thresholding of motions. RUC outputs are used to provide initial motions until storm tracking has stabilized, and improved methods for storm motion prediction include the ability to apply thresholds based on change in storm area. This helps eliminate false motion due to inaccuracies in the algorithm’s ability to define and track storms when the storm's area and shape significantly change. Finally NCWF-2 provides trending of dissipation based on grids as opposed to total storm area. This allows large storm systems to dissipate in one region and grow in another (Figure 4.3).


Figure 4.3. A 2-h NCWF-2 probability forecast. The green regions indicate storm observations based on radar and lightning data at valid time.

Standard CSI, POD, FAR, and Bias (Figure 4.4) were calculated based on binary grid-to-grid (no relaxation) comparisons for each probability level. In convective forecast verification statistics, the POD and CSI are often a reflection of the ratio of the forecast area to observation area (bias). This is evident in these performance statistics where the increased bias at low probability areas leads to higher PODs and CSIs.

Figure 4.4. Verification statistics compiled for 1-hr forecasts between 5 June and 15 August 2003. Each box plot represents verification for a different probability level (forecast 1 is >=10%, forecast
2 is >=20% etc.)

To better understand and calibrate the probabilities, reliability plots were compiled for various forecast periods using different filter sizes. Figure 4.5 shows an example reliability plot is shown for 0, 15, 30, 45, 60, 75, 90, 120, and 180 min NCWF-2 nowcasts from 10 July 2003. Relaxation was applied to the verification that allowed a margin around forecasts and observations of 10 km. The probabilities are calculated based on 2-km radar data and a 60x12-km elliptical filter. The 10 July case represents a mix of initiation, growth, and decay as well as various convective structures. The primary linear feature exhibits steady-state structure for about 6 h of the 24 h evaluation period.

Figure 4.5. Reliability chart for NCWF-2 at different forecast intervals. Forecast probabilities are along the x-axis and the percent of time observed along the y-axis. Perfectly-calibrated forecasts would lie along the straight black diagonal line. The forecast periods shown are 0 min (black), 15 min (red dashed), 30 min (green), 45 min (blue), 60 min (cyan), 75 min (magenta), 90 min (yellow), 120 min (gray) and 180 min
(dashed blue).

c) Use of RUC Analyses and Forecasts for Guidance for 0-3 h Nowcasts of Deep Convection.

S. Trier and A. Ahijevych (MMM) continued an ongoing collaboration with C. Mueller, D. Megenhardt and N. Rehak (RAP) to determine how RUC-2 analyses and forecast products can be used to assess the potential for growth and/or dissipation of mesoscale regions of deep convection over 0-3 h periods. Efforts this past year focused on combining, in a fuzzy logic framework, an objective algorithm that diagnoses the depth of thermodynamic instability weighted by other environmental conditions with RUC-2 3-h convective precipitation forecasts and a large-scale forcing field derived from RUC output.

This work began by examining how these elements were individually associated with the short-term evolution of deep convection for four test cases during summer 2002. The cases included meteorological situations in which mesoscale regions of thermodynamic instability were diagnosed, but the areal coverage and organization of deep convection differed. Combining an interest field based on large-scale forcing with the other interest fields, forecasts of mesoscale regions of organized convection were improved. At the same time, the technique reduced areal coverage of false alarms common in areas of thermodynamic instability that were previously removed from fronts or other linear regions of large-scale forcing. The new combined interest field was calculated during the FAA RCWF period during summer 2003. Point verifications confirmed that combinations of these fields provided better guidance for short-term convective forecasts than did use of any of the individual fields alone. Future work will examine how output from the combined interest field can be used to modulate short-term forecasts of convective growth based primarily on extrapolation of preexisting convection along a preferred orientation of strongest synoptic forcing.

5. Quantitative Precipitation Nowcasting

The goal of this USWRP-sponsored project is to enhance operational capabilities for quantitative precipitation nowcasting (QPN). During the past year activities took place in the following efforts.

a) Evaluation of the Sydney 2000 QPN efforts

A special issue of Weather and Forecasting is scheduled for February 2004 containing ten papers based on the Sydney 2000 Forecast Demonstration Project. RAP USWRP scientists were first authors on two of these papers (A. Crook and J. Sun [2004] and J. Wilson, et al. [2004]). and supporting authors on four other papers. This special issue helps establish the baseline capability for QPN and provide directions for future efforts.

b) Boundary-layer wind forecasts

A primary need for improved QPN is correct identification and forecast of the movement of boundary-layer convergence lines and characterization of their thermodynamic and kinematic properties. A major step was accomplished by parameterizing and incorporating evaporative cooling from precipitation into the dry version of VDRAS. This is accomplished by 4-D variational assimilation of radar data into VDRAS. The results clearly show improvement in forecasting the convergence and movement of gust fronts for a case from the STEPS project (Figure 5.1).

                 a) Analysis (initial conditions)                                  b) 1-hour forecast

Figure 5.1. Model 1-h forecast of gust front movement and evolution on 25 June 2002 from the Fort Wayne, IN NEXRAD radar. a) initialization time (1836 UTC) vectors are retrieved by model and overlaid on reflectivity. b) 1-hr model forecast (wind vectors) overlaid on 1936 UTC reflectivity field. White lines represent the leading edge of the gust front forecast by the model. The actual position can be seen as a reflectivity thin line.

c) Examination of new water vapor/stability measurement capabilities

The IHOP field project was conducted during 2002 partially for the purpose of determining if higher resolution measurements of water vapor would improve QPF. R. Roberts and J. Wilson are evaluating the use of visual, IR and reflectance satellite data and radar refractivity data to improve QPN. Roberts is working with J. Mecikalski and Feltz at the University of Wisconsin to obtain from satellite high-resolution stability fields for the IHOP period. A paper by R. Roberts and Rutledge (2003) was published during the year which describes a procedure for combining radar and satellite data to forecast convective storm initiation. The technique has been integrated into the AutoNowcaster.

J. Wilson and R. Roberts are also working closely with the Weckwerth Water Vapor USWRP/NCAR team to evaluate the quality of the S-Pol radar refractivity data to quantify near-surface water vapor. The radar-estimated water vapor is highly correlated with surface measurements, and with water vapor obtained from low-flying research aircraft. Results are reported in a paper by Pettet et al. (2003). Figure 5.2 is an example of the radar reflectivity and corresponding refractivity field for dry line case during IHOP.


     (a)                                                                         (b)

Figure 5.2. S-pol radar observation of a double-structured dry line on 22 May 2002. a) S-pol radar reflectivity field. The dry line (white arrow) appears as a thin line in the clear-air reflectivity field. A secondary dry line (black arrow) appears as a weaker thin line, b) S-pol radar refractivity field at the same time as the reflectivity image. There is a 2-3 gkg-1 water vapor difference across both the primary and secondary dry lines.

d) Origin and predictability of boundaries and predictability of convection.

An extensive study is underway by J. Wilson and R. Roberts to examine the initiation of all storms during IHOP. The predictability of QPN is closely tied to the predictability of the triggering mechanisms. All boundary-layer convergence lines and their influence on storm initiation and evolution are being identified and characterized. This includes characterizing the stability field within the region of storm initiation which makes use of special data sets acquired during IHOP. Based on these observations of storm initiation and triggering mechanisms the corresponding skill of numerical models to capture the initiating mechanisms and correctly forecasting precipitation initiation is examined. Two models are being examined, the RUC and MM5/LAPS. They were specifically run by FSL for IHOP. Preliminary results are reported by J. Wilson and R. Roberts, 2003 at the 31st Conference on Radar Meteorology. Figure 5.3 is an example of the type of analysis being conducted.






Figure 5.3. Examples of data and analysis being conducted for studies into the predictability of storm triggering mechanisms and predictability of storm evolution. a) color lines indicate analyzed convergence lines; (+) are storm initiation locations for a two-hour period centered on the time of boundary analysis (2100 UTC, 12 June 2003);b) boundary locations, surface station winds, composite reflectivity and storm initiation locations 0000 UTC, 13 June; c) same as b) for 0300 UTC, 13 June; d) CAPE from GOES sounder (blues 1000-2000 jkg-1, purples 0-1000 jkg-1) overlaid on visual satellite and analyzed boundaries at 2100 UTC, 12 June.

e) Temporal and spatial variability in atmospheric stability

R. Roberts and E. Nelson are examining several data sets collected during IHOP 2002 to address temporal and spatial variability in atmospheric stability and the effect on convective initiation. Objectives of this research are to document this variability and to demonstrate the resolution of data needed to improve forecasts of convective initiation. As a first step, S. Dettling developed a computer algorithm that ingests measurements from surface station networks (Figure 5.4) and sounding data collected by IHOP vertical profiling and rawinsonde systems. A Barnes analysis methodology was used to put the all of the station information onto a common grid spanning the IHOP domain.

The algorithm computes near-surface convergence as well as CAPE and CIN estimates at every mesonet station location every 5 min. The sounding closest to the station in distance and time (the user defines these limits) is used to update the sounding surface temperature (T) and dewpoint temperature (Td) values with their current mesonet values. The CAPE and CIN are updated using the new T and Td values. The resulting CAPE and CIN gridded fields are being examined in conjunction with hourly satellite sounder-derived CAPE and CIN fields (Figure 5.4-b), model-based stability fields and radar reflectivity images (Figure 5.4-c). Additional related research is discussed as part of the Water Cycle Strategic Initiative.


Figure (b) missing - placeholder   


Figure 5.4. Integrating IHOP 2002 data for addressing the variability in atmospheric stability as it relates to convection initaion. a) Convergence plot produced from a Barnes analysis of data from 18 different surface station networks on 12 June 2002 at 2200 UTC. b) CAPE field at 2200 UTC derived from GOES-8 satellite sounder data (courtesy of CIMSS) with satellite visible cloud field overlaid, c) Corresponding radar mosaic image over the same domain. Colored polylines represent convergence boundaries observed from radar, satellite and a subset of the surface stations shown.

f) IHOP Data Analysis

IHOP data are being used to study the details and predictability of convective storm initiation and evolution. The observations are being used to identify times and locations of storm initiation to determine the thermodynamic and kinematic mechanisms responsible for storm initiation and evolution. The predictability of storm initiation by numerical models and/or heuristic1 techniques depends on their ability to predict the initiation mechanisms. For example, if storms are initiated in a conditionally-unstable atmosphere from an updraft forced by the collision of two boundary-layer convergence lines, the predictability will be dependent on the ability to anticipate these phenomena.

Two IHOP periods with large mesoscale convective storm systems have been studied in detail. The study includes examining 3h and 6 h precipitation forecasts from a special 10-km grid version of the Rapid Update Cycle (RUC) model.

In both cases a weak synoptic-scale low and associated synoptic fronts were present. While the storms initially developed along the fronts or near the low pressure area, evolution to the mesoscale convective storm system phase was controlled by individual gust fronts. Characteristics of the gust fronts, such as the magnitude of the associated convergence and relative motion between the storms and gust front, were controlling factors for storm evolution. In each case the RUC forecasted the initial storms in about the correct locations. However, the timing of the initiations, evolution and motion of subsequent convection was incorrect. This is likely due to the inability of the numerical models to correctly forecast which storms will produce gust fronts and the characteristics of these gust fronts.

These preliminary findings have important implications for modeling activities. If numerical models are to correctly forecast the evolution and propagation of convective storms they need to be able to forecast the evolution and characteristics of the associated gust fronts. The magnitude of the outflows from thunderstorms is heavily dependent on the characteristics of precipitation in thunderstorm downdrafts. A promising area for research is to use polarimetric radar to estimate these microphysical parameters and assimilate them into explicit storm numerical forecast models.


1Heuristic is defined here as forecast rules based on experiment, numerical simulations, theory and forecaster rules of thumb.