Johannesburg Lightning Nowcasting From Meteorological Data and Electric Field Using Machine Learning

Abstract

This paper presents the work and findings of using machine learning algorithms namely, logistic regression (LR), random forest (RF), and long short-term memory (LSTM) network, to nowcast Johannesburg's cloud-to-ground lightning events within a 30 km radius of the city center between the period of 1 November 2021 to 27 February 2022. The investigation evaluated each model's ability to nowcast lightning strokes over the city from recorded historical values of the electric field, air temperature, dew point, and relative humidity for a forecast horizon of 15 minutes. Performance metrics indicate that the recall score for the LSTM is the lowest at 53% while that of the RF and LR are 80% and 93% respectively. The RF and LSTM models achieved lower recall scores but demonstrated less sensitivity to making false predictions, while the logistic regression made a relatively higher number of false positive misclassifications. A precision score of 41% for the LSTM indicates that the model is able to predict non-lightning occurrence more precisely than the LR and RF models which reported precision scores of 9% and 11% respectively. The models' prediction performance over the Sentech and Hillbrow towers has also been assessed and analyzed, and results indicate that a model's predictive ability is heavily influenced by cloud-to-cloud lightning.