Short-term Prediction of Lightning in Southern Africa using Autoregressive Machine Learning Techniques

Authors:

Abstract
Lightning is responsible for both human and economic loss but its prediction remains challenging. We seek to find a lightning prediction model in South Africa that uses historical lightning-flash data only. This type of prediction model is cost-effective, easy to interpret and may be used for realtime forecasting. We evaluated and compared three popular time-series machine learning techniques on their ability to predict the number of Cloud-to-ground lightning flashes in South Africa for three-hours ahead. These models are the Auto Regressive (AR), Auto Regressive Integrated Moving Average (ARIMA) and the Long-Short-Term-Memory Recurrent Neural Network (LSTM) models. We used historical lightning data from the South African Lightning Detection Network during 2018. Our prediction model parameters were AR(lag=8), ARIMA (AR lag=8, Integrate=0, MA lag=2) and LSTM (2x50 layers, activation=ReLU, optimizer=adam) and models were minimized for Root Mean Square Error but evaluated based on Mean Absolute Percentage Error (MAPE). We used a 70%/30% Train-test split. The AR and ARIMA models performed comparably with a MAPE of 15312 and 15080 respectively. The LSTM Model outperformed considerably with a MAPE of 3705. Although the LSTM model outperformed, predictions errors in absolute terms were still high. This paper highlights the usefulness of nonparametric predictions models for lightning prediction.