Weather, especially convective storms, exerts a disruptive influence on aviation, both in the terminal area and en-route air traffic flow. For daily planning purposes, aviation users need forecasts with outlook horizons several hours into the future that provide not only details about the likely weather outcome, but also information about storm structure, intensity, and organization, and the associated forecast uncertainty. Weather predictions are inherently uncertain and this prediction uncertainty increases with increasing forecast time horizon. The numerical weather prediction community is moving toward utilization of ensemble approaches to characterize prediction uncertainty and expressing forecasts in probabilistic terms.

The NCAR Research Applications Laboratory (RAL) has been developing new concepts of how probabilistic weather forecasts can be tailored for aviation needs. The focus of this research is to develop tools that distill large amounts of observational and ensemble model data into a probabilistic prediction of convective storms with attributes that are likely to impact air traffic flows across the NAS. In this preliminary effort, the hazard being addressed is defined as contiguous areas of VIP 3+ exceeding 100 km in length and lasting longer than 1 hour. It should be noted that this definition is being used for proof of concept and that the definition of the aviation hazard may be refined to include echo tops, lightning flash rates, and minimum area constraints, as deemed appropriate.

This product is designed to provide an estimate of the likelihood that Large-scale Convective Storms (LCS) may be present or form in the next 1-23 hours. Current development uses both the hourly-updating NOAA/GSD High Resolution Rapid Refresh (HRRR) and the AFWA multi-model ensemble to calculate LCS and LCS Initiation likelihood forecasts. Convective areas are defined as those with radar indicated VIL exceeding 3.5 kg m-2 (i.e., VIP Level 3+) in the observations. A calibration prodecure found that the optimal VIL threshold to use to detect LCS in the HRRR data was 2.5 kg m-2 to compensate for HRRR's tendency to underforecast VIL. LCS are indicated when the maximum dimension of this area of VIL (allowing for gaps of up 10 km) exceeds 100 km for more than 1 hour. LCS CI is detected by searching back 2 hours in time and within a 125 km radius for any previously existing LCS. If there LCS are present in the past 2 hrs and 125 km, the LCS is considered an initiation event.

The HRRR model forecast data from all currently valid forecasts are used to generate the time-lagged ensemble forecast. As a simple example, consider a 2h HRRR forecast generated at 00 UTC and a 1h forecast generated at 01 UTC. These two runs comprise a 2-member time-lagged ensemble valid at 02 UTC. For our purposes, we require least 5 valid forecast leads to be available to compute the likelihood of and LCS or LCS-I event for a given leadtime. forecast. The number of members available varies depending on the latency of the HRRR (which was typically about 2 hours) and the length of HRRR runs generate in the last 15 hours. A new time-lagged ensemble based LCS likelihood forecast is generated each hour, regardless of whether or not a new HRRR forecast is available.

The AFWA multi-model ensemble is a high resolution model that uses varying physics and boundary forcing to generate a 10 member ensemble. This ensemble is updated twice per day (00 and 12 utc) with a forecast period of more than 36 hours. Only the first 30 hours of this forecast are used to compute the likelihood of LCS and LCS-I. Based on analyses this spring an optimized threshold of 5 kg m-2 is used to identify LCS in each member of the multi-model ensemble.

A couple of other constraints are placed on the calculation of the LCS likelihood fields. Data from lead hours 0-2 are not used due to model spin-up. At least 5 forecasts must be available for a given forecast valid time in order to calculate the likelihood field. A distance weighted spatial filter of 125 km radius is used to extend the region for potential LCS activity to account for possible position errors in the HRRR forecasts with weight decreasing with distance from the model forecasted storm location.

This work has been supported by the Federal Aviation Administration (FAA). Any opinions, findings, and conclusions or recommendations expressed on these websites are those of the authors and do not necessarily reflect the views of the sponsoring agencies.

Prototype data can be viewed using a java/javaws based application developed at NCAR/RAL called JAZZ.

Click here --> JAZZ LCS Data Viewer <-- to start up java data viewer application.

In order for JAZZ to work you will nee to have java with javaws installed on your machine

- Tested for linux os using java version 1.7.0_25
- typical path --> /usr/local/jdk1.7.0_25/bin/javaws

There are 7 products available for viewing in this display under the "Grids" Tab:

- OBS NSSL VIL - mosaic of vertically integerated liquid water disseminated by NSSL.
- OBS LCS Detection - Large Convective Storm (LCS) detections that depict the shape and size of the storm in red with a 125 km buffer around each storm in yellow. There is about a 1 hour delay in the detection of LCS in the observations due to the requirement that storms persist for at least 1 hour.
- OBS LCS CI Detection - Large Convective Storm (LCS) Convective Initiation detections that depict the shape and size of the storm in red with a 125 km buffer around each storm in yellow. There is about a 1 hour delay in the detection of LCS CI in the observations due to the requirement that storms persist for at least 1 hour.
- HRRR LCS Likelihood* - interest field indicating the likelihood of LCS being within 125 km and 1 hour of the valid time based on HRRR time-lagged ensemble with forecasts extending up to 10 hours into the future.
- HRRR LCS CI Likelihood* - interest field indicating the likelihood of LCS initiating within 125 km and 1 hour of a point in time and space based on HRRR time-lagged ensemble.
- AFWA LCS Likelihood* - interest field indicating the likelihood of LCS being within 125 km and 1 hour of the valid time based on AFWA 10 member multimodel ensemble with forecasts extending up to 24 hours into the future.
- AFWA LCS CI Likelihood* - interest field indicating the likelihood of LCS initiating within 125 km and 1 hour of a point in time and space based on AFWA 10 member multimodel ensemble. The following 2 products are currently unavailable:
- AFWA Prob of 50% East-West Capacity Reduction at 30 kFT - Probability that at least 50% of the capacity for east-west air traffic flows will be lost at a scale of 100 km at 30 kFT based on 10 member AFWA high-res (4km) muti-model ensemble.
- AFWA Prob of 50% North-South Capacity Reduction at 30 kFT - Probability that at least 50% of the capacity for north-south air traffic flows will be lost at a scale of 100 km at 30 kFT based on 10 member AFWA high-res (4km) muti-model ensemble.

*note: The interest field actually ranges from 0-2 and is not currently cast as a probability, however, higher values can be interpreted of increased likelihood. We are in the process of collecting data to develop a calibrated probablities.

Pulldown Tabs at top of the page can be used to select data fields and to perform minimal configuration of the display.

- * File - export an image
- * Grids - select data field for display * Features - not available
- * Maps - toggle map overlays
- * View - zoom in on specified region
- * Configure - adjust properties of display and how grids are displayed
- - Grid - adjust grid transparency, line contours
- - Layer - adjust order of layering - obs data typically best for top layer
- - Time - adjust time bar properties

Buttons below Pulldown Tabs - mouse over for explanation

Time bar features

- "Now" button - updates data in viewer to latest available - current time shifts to rightmost location on timebar
- "Time-bar"
- * Right-click on time bar to select a specific time of interest.
- * Right-click and Drag on arrow heads above timebar to the left or right to "stretch" the time bar to extend further into the past or future. Note - This will need to be done if you are interested in forecast data and have recently used the "Now" button.
- "Movie buttons" - allow you to loop through products forward, backward or sweep mode

The performance of the deterministic, hourly updating NOAA/GSD HRRR model is assessed by evaluating the skill of its predictions of LCS and LCS-I events.

- The detection of LCS and LCS-I events in the model VIL field has been optimized using a dynamically-updating Iterative Threshold Optimization Algorithm (IOTA)
- The skill of the model at predicting LCS and the dynamic thresholds used to detect LCS are provided.
- The time-lagged ensemble of calibrated LCS detections is used to produce reliable forecasts of the likelihood of LCS occuring within 50 km of a given location.

(1) | Num. of storms / day (obs, HRRR) | LCSs plot 1 LCSs plot 2 | LCS-Is plot 1 LCS-Is plot 2 | ||||||

(2) | HRRR LCS prediction performance stats (all) | Stats eUS GP MRV APP TX SE | Stats eUS GP MRV APP TX SE | ||||||

(3) | HRRR LCS stats during past 30 days | PctMiss eUS GP MRV APP TX SE | FAR eUS GP MRV APP TX SE | FAratio eUS GP MRV APP TX SE | PODno eUS GP MRV APP TX SE | CSI eUS GP MRV APP TX SE | SEDS eUS GP MRV APP TX SE | Bias eUS GP MRV APP TX SE | |

(4) | HRRR LCS-I stats during past 30 days | PODno, CSI, SEDS eUS GP MRV APP TX SE | Bias, FA Ratio eUS GP MRV APP TX SE | ||||||

(5) | HRRR LCS-I prediction performance stats | Stats eUS GP MRV APP TX SE | Stats eUS GP MRV APP TX SE |

- row 1 - LCS and LCS Initiation storms counts - number of modeled (HRRR & AFWA) and observed events matched in time only - last 10 days
- row 2 - HRRR 7-day history of stats for LCS matched in space (within 1 deg) and time (no time window) as a function of leadtime for eastern 2/3 of U.S. (all issue times combined)
- row 3 - HRRR 30-day history of stats for LCS matched in space (within 1 deg) and time (no time window) as a function of leadtime for eastern 2/3 of U.S.
- row 4 - HRRR 30-day running mean stats for LCS_I matched in space (within 2.5 deg) and time (no time window) as a function of day and leadtime for eastern 2/3 of U.S.
- row 5 - HRRR 30-day mean stats for LCS_I matched in space (within 2.5 deg) and within 0,+/-1 & +/-2 hrs as a function of leadtime for eastern 2/3 of U.S.

- ROC Area versus leadtime - area under ROC (Receiver Operating Characteristic) curves - higher values are better
- Reliability Diagrams - observed frequency of occurrence for a given forecast category - want to be close to 1:1 line
- above 1:1 line = forecast likelihood tends to under-forecast observed frequency
- under 1:1 line = forecast likelihood tends to over-forecast observed frequency

*Note: statistics are for CONUS east of 105W and do not include data from Canada or coastal waters

The LCS and LCS Initation detection capabilities are applied to model ensemble data to :

1) Provide estimates of the likelihood that convection or convection initiation with a max dimension exceeding 100 km will occur at a certain time and within a distance of 125 km of a given grid point.

2) Provide a measure of model's ability to predict organized convection to aid aviation weather forecasters in estimating confidence in the prediction of large-scale storm initiation or occurrence.

** An LCS is defined as a storm exceeding 100 km in length (including gaps of less than 10 km) and persisting for more than 1 hour.

** An LCS Initiation is an LCS that occurs in a region that is at least 125 km away from any LCS that existed in the last 2 hours.

** The forecasts are currently cast as likelihoods since they have not yet been calibrated. Reliability diagrams for HRRRindicate that the HRRR LCS likelihoods are a good estimate of probability up to a likelihood value of around 0.5. Higher likelihood values tend to overforecast the true probability. Currently, the HRRR LCS-I forecasts tend to strongly overforecast the true probability. Similar trends can be seend for the AFWA LCS and LCS-I likelihood forecasts.

Each individual ensemble member is evaluated in its ability to predict LCS and LCS-I.

- Comparing model and obs LCS and LCS-I counts (unmatched in space) gives an indication of member-by-member biases
- Stats vs leadtime (matched in space and time) - generally skill decreases with forecast leadtime by about 40% in first 6 hours of HRRR run
- Month-long stats plots (matched in space and time) - show trends in the performance of each leadtime over the last 30 days.
- The ability to predict an LCS Initiation event (which occurs at a single instance in time) is evaluated using a varying time window (up to +/- 2 hours) and a much larger allowable distance buffer (2.5 deg) The variation in skill with time-window size can be used to assess how often forecast is within 1 and 2 hours of the actual event.

Data are harvested from high resolution ensemble models (HRRR and AFWA). A great deal of information is distilled into a single forecast that provides a quicklook display of large-scale storm forecasts that can be used to assess where problem areas are most likely to occurr in the next 2-24 hours.

Because of its rapid update cycle (hourly), the HRRR model forecasts are used to generate a time-lagged ensemble. The HRRR has the advantage (over AFWA) of assimilating 3D radar reflectivity and updating hourly. AFWA is a multli-model ensemble that updates twice daily. Multimodel ensembles generally have better spread than that obtained from a time-lagged ensemble. Thus, offering a better depiction of forecast uncertainty (and the range of possible outcomes) than the HRRR. The AFWA LCS likelihoods provide a longe range outlook of where and when LCSs may exist in the next 6-24 hours while the HRRR LCS likelihoods (which are more accurate and more reliable than AFWA out to 8 hours) can be used to refine the longer-range AFWA predictions.

Examples of potential usage:

- The LCS likelihood field can be used to aid in estimating forecast confidence.
- The LCS_I likelihood field can be used to identify where new problem areas are possible that forecasters can focus additional resources on to refine likelihoods, location and timing of LCS initiation events.