Thunderstorm Forecasts

Short Term Forecast

South Baldy Extended Forecast

South Baldy Thunderstorm Forecast Methodology

Short Term Forecasts

Short term (i.e., 9-to-33-hr) thunderstorm forecasts for use at the Langmuir Laboratory for Atmospheric Research.  Forecasts will be issued by me (and occasionally a student trainee meteorologist for whom I am training), Sunday through Thursday at 9pm Mountain Time, and will be valid for the following day. These forecasts will be based on top-of-the-hour (e.g., 00, 01, 02Z), time-weighted output (that lends more credence to cycles nearer to times forecast) from the 18Z and 00Z cycles of two convective-allowing, numerical weather prediction models; namely the High Resolution Rapid Refresh (HRRR) model, and the 3-km North American Mesoscale Model (3-km NAM). These models output simulated composite radar reflectivity (SCR) and thermodynamic properties of the environments for which thunderstorm activity is situated[1]. 

To forecast thunderstorm activity, I will be using the following conditions: 1) an SCR threshold greater than 29 dBZ to determine whether heavy, presumably convective, precipitation is forecast; 2) a lifted condensation level temperature (LCLT) greater than -11 °C; and, 3) a Cloud Physics Thunder Parameter (CPTP) greater than zero [dimensionless] (see Bright et al (2005); and see Stull (2017) p. 569). The LCLTs and CPTPs will be from 1-hr before SCR output in an attempt to limit convective contamination. This seems reasonable since a Byers-Braham (1949) cell (see Stull (2017) p. 484-85) lasts on the order of 30-to-60-min. The LCLT and CPTP will be calculated from the model gridcell with the greatest SCR within 20-km (easting and northing) of South Baldy (since the 20-km range overlap with multiple model gridcells.) 

Since there are two models and two cycles of each model at every forecast time though, whether lightning is forecast to occur will depend on the sum of the time-weighted assignments (i.e., 0 or 1 times the weight) for every forecast time. Note: it is probably true that there is no "magic number" when it comes to weighting numerical model output; however, given the same or similar data streams (assuming no errors in initial conditions) and resolutions it is reasonable to assert that model cycles nearer to the time of the event (i.e., lightning) should provide more accurate (and precise) forecasts. With that in mind, my "time-weighted" scheme simply accounts for the time difference between model cycles; such that for longer range forecasts, the difference in time between one model cycle and the next becomes increasingly small relative to the forecast time.  

Forecast periods are 6-hr in length. If any forecast time (e.g., 15Z, 16Z, or 17Z over the 12-18Z period) within a 6-hr period is favorable for lightning, that (entire) period will be regarded as being favorable for lightning and will be assigned a "YES" in the table above under "Short Term Forecast," accordingly. Additionally, probabilities from one time to the next within a 6-hr forecast period combine; in accordance with the binomial distribution[2]. In cases where conditions are favorable for lightning at the start of a 6-hr period, the last hour of the preceding (entire) 6-hr period will be regarded as being favorable for lightning (since deep, precipitating convection does not develop instantaneously in these models). 

That said, forecasts for thunderstorm activity within 20-km (northing and easting) of South Baldy will be conveyed as a "YES" or "NO" for each 6-hr period in the (simple) table above under "Short Term Forecast" as well as in once-daily email messages to Langmuir personnel (that will include a "Short Term Forecast" discussion for those who want more detail). Verification of these forecasts will be done using data from the Langmuir Lightning Mapping Array[4]. Occasionally, I will post verification style graphs in figure plots under the "South Baldy Thunderstorm Forecast Verification" section below just to see how these forecasts are panning out (and as a means to improve these forecasts).

Extended Forecasts

Extended (i.e., 33-to-105-hr) thunderstorm forecasts for use at the Langmuir Laboratory or Atmospheric Research.  Forecasts will be issued by me (and occasionally by a student trainee meteorologist for whom I am training), Sunday through Thursday at 9pm Mountain Time, and will be valid for 2-to-4 days out. These forecasts will be based on top-of-the-hour, space-weighted output (that lends more credence to the higher resolution model of the two in terms of gridcell spacing) from the 18Z cycle of two medium range numerical weather prediction models: the Global Forecast System (GFS) model, which runs at approximately 28-km (or exactly 1-Decimal Degree (DD) in terms of longitude and latitude) horizontal resolution out to 384-hr from the time of initialization; and, the North American Mesoscale Model (NAM), which runs at 12-km horizontal resolution out to 84-hr from the time of initialization. Similar to HRRR and 3-km NAM, the GFS and NAM models output SCR and thermodynamic properties of the environments for which thunderstorm activity is situated[1]. However, in contrast to HRRR and 3-km NAM, the GFS and NAM are coarser resolution with respect to space and time and are poorer at being able to capture the space and time occupied by (individual) Byers-Braham (1949) cells (see Stull (2017) p. 484-85). As such, neither the GFS or NAM are convective-allowing (which is typical for medium or longer range models because of the coarser resolutions). So, instead of output being available every hour at 3-km (horizontal) resolution as is the case with the HRRR and 3-km NAM, the GFS and NAM are only available at 28-km and 12-km (horizontal) resolution, respectively, every three-hours (e.g., 00, 03, 06Z, and so on). Additionally, since the aim is for a 2-to-4 day forecast here, it is worth noting that only one of these models, namely the GFS, offer solutions to cover later times in the forecast beyond 84-hr from the time of initialization. So, beyond 84-hr from the time of model initialization, only the GFS is available for forecasts. Nevertheless, the model time steps for GFS and NAM are identical and each model has a convective parameterization scheme that can be used to identify deep, precipitating convection. 

To forecast thunderstorm activity, I have chosen to use an algorithm similar to what I use for short term forecasts: 1) an SCR threshold greater than 29 dBZ to determine whether heavy, convective, precipitation is forecast; 2) LCLT greater than -11 °C; and, 3) a CPTP greater than zero [dimensionless] (again, see Bright et al 2005; and see Stull (2017) p. 569). However, in contrast to short term forecasts, the LCLTs and CPTPs will now be from 3-hr before SCR output in an attempt to limit convective contamination. While I am have some valid concerns regarding this long of a leap in time (i.e., 3-hr), mainly because, again, a (typical) Byers-Braham (1949) cell (only) lasts on the order of 30-to-60-min, I am not convinced that the alternative is much better. This is because the "alternative" would involve using a best-fit line to interpolate dewpoint and temperature up to 1-hr in advance of model storm activity using output valid from 2-hr prior; effectively generating a 2-hr forecast of dewpoint and temperature on top of what is already being forecast (i.e., lightning). That said, LCLT and CPTP will be calculated from the model gridcell with the greatest SCR within 20-km (easting and northing) of South Baldy, similar to short term forecasts. 

As previously indicated, since the GFS model run at 28-km horizontal resolution and the NAM model run at 12-km horizontal resolution, there are approximately 5.4 times as many NAM gridcells as GFS gridcells within the domain. So, given these differences in terms of the number of gridcells between GFS and NAM, whether lightning is forecast to occur will depend on the sum of the space-weighted assignments (i.e., 0 or 1 times the weight) for every forecast time wherever there is overlap.  My "space-weighted" scheme weighs NAM output 5.4 times as much as GFS output for any given hour. Again, because there are 5.4 times as many NAM gridcells as GFS gridcells in the domain, and because I am trying to resolve something (i.e., a thunderstorm) that is finer in scale than the (horizontal) space resolution of each model. However, since the GFS offer solutions farther out than NAM, all of the weight is shifted to the GFS beyond 84-hr from the time of model initialization (i.e., 18Z, everyday) since the NAM only runs out to 84-hr from the time of its initialization. That said, note: similar to the time-weight scheme used in short term forecasts, it is probably true that there is no "magic number" here either when it comes to weighting numerical model output; however, given the same or similar data streams (assuming no errors in initial conditions), it is reasonable to assert that the model nearer to the scale of a thunderstorm in this case should provide more accurate (and precise) forecasts of a thunderstorm. 

Forecast periods are 12-hr in length versus the 6-hr in length periods used in my short term forecasts. I decreased the resolution of the forecast periods to roughly match the decreased resolutions of the models being used in extended forecasts relative to short term forecasts. Activities at Langmuir Laboratory also tend to be grouped as being either "daytime" or "nighttime," so use of a 12-hr period seems appropriate if the beginning and end of periods are set to be within a few hours of sunrise and sunset (i.e., 6am and 6pm, Mountain Time). That said, if any forecast time (e.g., 12Z, 15Z, 18Z, or 21Z over the 12-00Z period) within a 12-hr period is favorable for lightning, that (entire) period will be regarded as being favorable for lightning and will be assigned a "YES" in the table above under "Extended Forecast," accordingly. Additionally, probabilities from one time to the next within a 12-hr forecast period combine; in accordance with the binomial distribution[3]. In cases where conditions are favorable for lightning at the start of a 12-hr period, the 3-hr preceding the (entire) 12-hr period will be regarded as being favorable for lightning (since deep, precipitating convection does not develop instantaneously in these models, even though it is not explicitly resolved in the case). 

That said, forecasts for thunderstorm activity within 20-km (northing and easting) of South Baldy will be conveyed as a "YES" or "NO" for each 12-hr period in the (simple) table above under "Extended Forecast" as well as in once-daily email messages to Langmuir personnel (that will include a "Extended Forecast" discussion for those who want more detail).  Verification of these forecasts will be done using data from the Langmuir Lightning Mapping Array[5]. Occasionally, I will post verification style graphs in figure plots to "Extended Forecasts" under the "South Baldy Thunderstorm Forecast Verification" section below just to see how these forecasts are panning out (and as a means for improving forecasts).

South Baldy Thunderstorm Forecast Verification

Short Term Forecasts

Figure 1. Forecast and Observed Agree with respect to Period: Morning 98% Accurate (Out of 196 Forecasts), Afternoon 92% Accurate (Out of 198 Forecasts), Evening 94% Accurate (Out of 197 Forecasts) and Overnight 94% Accurate (Out of 198 Forecasts). June 9, 2024-through-March 31, 2025. Data Verified using the Langmuir Lightning Mapping Array[4].

Figure 2. True Positive with respect to Period. This is when a Thunderstorm was Forecast and a Thunderstorm was Observed: Morning 100% Accurate (Out of 1 Forecast), Afternoon 92% Accurate (Out of 37 Forecasts), Evening 72% Accurate (Out of 29 Forecasts) and Overnight 18% Accurate (Out of 11 Forecasts). June 9, 2024-through-March 31, 2025. Data Verified using the Langmuir Lightning Mapping Array[4].

Figure 3. True Negative with respect to Period. This is when a Thunderstorm was Not Forecast and a Thunderstorm was Not Observed: Morning 100% Accurate (Out of 191 Forecasts), Afternoon 98% Accurate (Out of 151 Forecasts), Evening 95% Accurate (Out of 174 Forecasts) and Overnight 95% Accurate (Out of 194 Forecasts). June 9, 2024-through-March 31, 2025. Data Verified using the Langmuir Lightning Mapping Array[4].

Figure 4. False Positive with respect to Period. This is when a Thunderstorm was Forecast and a Thunderstorm was Not Observed (i.e., Over-Forecast): Morning 83% Inaccurate (Out of 6 Forecasts), Afternoon 28% Inaccurate (Out of 47 Forecasts), Evening 9% Inaccurate (Out of 23 Forecasts) and Overnight 50% Inaccurate (Out of 4 Forecasts). June 9, 2024-through-March 31, 2025. Data Verified using the Langmuir Lightning Mapping Array[4].

Figure 5. False Negative with respect to Period. This is when a Thunderstorm was Not Forecast and a Thunderstorm was Observed (i.e., Under-Forecast): Morning 0% Inaccurate (Out of 200 Forecasts), Afternoon 2% Inaccurate (Out of 157 Forecasts), Evening 4% Inaccurate (Out of 183 Forecasts) and Overnight 4% Inaccurate (Out of 203 Forecasts). June 9, 2024-through-March 31, 2025. Data Verified using the Langmuir Lightning Mapping Array[4].