Atmospheric Meteorological Effects on Forecasting Daily Lightning Occurrence at Cape Canaveral Space Force Station

As the Cape Canaveral Space Force Station and Kennedy Space Center increase their launch rate, any process that could assist in the automation of the currently-manual lightning forecast would be valuable. This work examines the possibility of machine-learning assistance with the daily lighting forecast which is produced by the 45th Weather Squadron. A dataset consisting of 34 lightning, pressure, temperature and windspeed measurements taken from 334 daily weather balloon (rawinsonde) launches in the timeframe 2012–2021 was examined. Models were created using recursive feature elimination on logistic regression and XGClassifier algorithms, as well as Bayesian and bandit optimization of neural network (NN) hyperparameters. The modeling process was repeated after eliminating 13 features related to windspeed. The best performing models on both datasets were the optimized NN models, with an F1 metric of 0.79 on the full dataset and 0.66 on the reduced dataset. For comparison, a model that predicted randomly achieved F1 = 0.47 on this dataset. The addition of 13 windspeed-related features more than doubled the complexity of the 21-feature no-wind model while increasing model performance by 13 percentage points. A notable inference from the statistical modeling is that the most important feature from both datasets was the Thompson convective index, which is related to temperature, dewpoint, relative humidity and lapse rate.