Lightning Prediction with Deep Learning at Varying Temporal and Spatial Scales

Abstract

Lightning directly causes human fatalities, sparks costly wildfires, and can be triggered during rocket launches, potentially harming high-value payloads. This threat is increasingly consequential as launch activity at Kennedy Space Center exceeds 100 launches per year. Observing the microphysical processes that trigger lightning is difficult. Recent advances in radar technology have allowed for charged regions of storms to be quantified using specific differential phase (KDP ), at temporal scales of seconds with a fully digital phased array radar. I developed three deep learning models (BoltCast, LaunchCast, and HorusCast), investigating machine learning’s ability to generate lightning probabilities or its related dual-polarization variable, KDP . BoltCast’s domain spans the entire CONUS with a four-day prediction, LaunchCast is near Kennedy Space Center along the east-central Florida coastline with a 1-hour prediction, and HorusCast translates all traditional and dual-polarization variables into a KDP estimate. Each model was developed using custom deep neural networks. The models are evaluated with reliability or attributes diagrams, precision-recall performance diagrams, or error calculations; with varying explainability metrics applied to identify the main drivers of the model. BoltCast was evaluated against two separate model architectures and showed high reliability out to four days, with area under the precision-recall curve (AUC-PR) values decreasing from 0.62 to 0.52, for the best architecture. Permutation importance showed that CAPE was the primary driver, with partial dependence plots showing larger changes in model output when CAPE and reflectivity are varied for a hit and miss forecast. LaunchCast transitions to a much smaller scale, making lightning predictions out to 1-hour, at 15-minute time steps, with 5 High Resolution Rapid Refresh variables, a subset of the Multi-Radar Multi-Sensor observation suite, Geostationary Lightning Mapper group area, and electric-field-mill data as inputs. AUC-PR drops from 0.47 to 0.26 across the forecast hour. The accumulated local effect of a subset of input data revealed that the mean electric-field-mill values have influence within 5 nautical miles of the site. Lastly, HorusCast, a KDP estimator, generates fairly strong KDP predictions when the radar is configured to scan continuously. The model performance drops when deployed on test cases that have different scanning strategies. Input x Gradient (IxG) revealed that the differential phase ΦDP had the greatest attribution. This work demonstrates the value of deep-learning’s application for lightning specific predictions and KDP.