VLab Forum Members,
Just a quick reminder that the June 2023 VLab Forum will occur this
Thursday, the 22nd, at 3:35 PM – 4:30 PM (Eastern Time). The talk
features a presentation titled "Probabilistic Forecast of
Thunderstorms Using 2-D Convolutional Neural Network (CNN)"
which is being presented by Mamoudou Ba. We hope you can attend.
To participate in the forum, please register
for the webinar.
Abstract:
A supervised 2-Dimensional (2-D) Convolutional Neural Network (CNN)
model was developed by training the model with 5 years (28 September
– 31 October 2016, and March 1 – 31 October 2017 - 2020) of
predictors derived from the High-Resolution Rapid Refresh (HRRR)
model to produce short range (1 - 8h) CNN Probability of
Thunderstorm Forecasts (CNNPTFs) over the conterminous United States
(CONUS). An architecture of three hidden layers of 80 features
detection maps each is used. Thirteen HRRR variables that provided
the most accurate predictions among 22 predictors experimented with
during the development of the algorithm were selected.
Multi-Radar/Multi-Sensor System (MRMS) Composite Reflectivity
(MRMSCRs) values were used as truth data to train the model. CNNPTFs
were produced over our validation period (1 March - 31 October 2021
and 2022) at each HRRR model’s initiation time. Qualitative
validation was performed by comparing CNNPTFs to MRMSCRs for several
case studies during the validation period. Finally, traditional
statistical metrics are computed to objectively verify the skill of
CNNPTFs and compare it performance to that of the HRRR Simulated
Composite Reflectivity Forecasts (HRRRSCRFs). Compared to HRRRSCRFs
over our validation period the performance of CNNPTFs generally
exceeds that of HRRRSCRFs.
In this presentation, the CNN technique will be discussed, the
methodology will be presented, and results showing the skill of the
CNN will be shown. While this research is being undertaken with the
goal of improving MDL’s thunderstorm forecast guidance, the AI CNN
technique has potential application to other post-processing work
being done in MDL.
Agenda:
You can find the agenda for the Forum at the following link:
Add to Your Calendar:
To add this VLab Forum meeting to
your calendar, please click on the following button.
