Lightning Prediction at Cape Canaveral AFS, FL, via the Application of Deep Learning to Surface Observations and Field Mill Data

Cape Canaveral Air Force Station (CCAFS), Kennedy Space Center (KSC), and Patrick Air Force Base (PAFB) all reside in the thunderstorm capital of the United States. These installations experience more thunderstorms per year than any other place in the world (Florida Climate Center, 200). Therefore, lightning poses a frequent and significant risk to the over 25,000 people work who in and around the $17+ billion launch facilities to assemble billions of dollars in rocket boosters and their associated payloads (Roeder, Hajek, Flinn, Maul, & Fitzpatrick, 2000). It is the mission of the 45th Weather Squadron (45th WS) to provide timely and accurate warnings of weather conditions such as lightning that pose a risk to assets and personnel CCAFS, KSC and PAFB.
To aid 45th WS forecasters, a network of 30 Electric Field Mills (EFMs) was installed in the area in and around CCAFS, KSC and PAFB. EFMs record the electrification of the local atmosphere. Several efforts have been made over the years to find an optimal way to utilize the EFM network data to improve lightning prediction. These efforts approached the problem using atmospheric science as well as traditional statistical regression techniques with mixed results. The most recent effort by Hill (2018) trained a Convolutional Neural Network (CNN) on EFM network data from the months of May-July of 2012-2016. The mean was calculated for every 60 second period. Then, 30 minutes of this summarized data was used to predict lightning with a warning period of 15 minutes. This method achieved accuracy and precision rivaling current best practices for lightning prediction. This suggested that the EFM sensor array may contain some portion of the total information required to effectively predict lightning for a 5-mile radius near CCAFS.

In this paper, hourly statistics were generated from the raw EFM data set used in Hill (2018). Input variables were generated from surface observations from every station within 50 miles of CCAFS and then combined with the EFM statistics for the same time periods. Further variables were added to the data set to capture the current trend of surface observations and EFM readings. This combined data set was then treated as if each entry were an image in a sequence of images where each variable value was a pixel color value. A set number of “images” were then selected to make up a sequence. The resulting data set of “image” sequences was then fed into a deep learning CNN Long Short-Term Memory (CNN-LSTM) network in order to predict whether future time periods would fall into the “lightning” or “no lightning” category. Metrics are provided on performance of the CNN-LSTM as well as insights into other methods of further improvement at lightning prediction at CCAFS using deep learning.