A LSTM Recurrent Neural Network for Lightning Flash Prediction within Southern Africa using Historical Time-series Data

We evaluated the prediction ability of a Long-Short-Term-Memory Recurrent Neural Network (LSTM) model to predict short-term lightning flash density within South Africa using historical lightning events. We predicted the lightning flash densities for one-hour, three-hour and 24-hour periods for two areas within Southern Africa using data from the South African Lightning Detection Network. These areas represent a moderate and a high-flash density area. The inputs were (1) historical one-hour lightning density, and (2) a five-day lightning density average for the corresponding period in the previous year. Models were trained using four years of data and predictions were made for every one-hour interval for one-year. Our LSTM model is composed of 1×50 network layer, 2×25 layers, 1 dense layer (activation=Leaky ReLU, optimizer=adam). Models were minimized for Mean Square Error (MSE) but evaluated based on Mean Absolute Error (MAE) and MSE. Our models were tested for repeatability by re-running the model for prediction years 2012, 2013, 2014 and 2015. The results showed a combined MAE of 2.87 flashes.hr-1 and combined MSE of 1209 for one-hour predictions. MAE values remained relatively similar but standard deviations of the errors increased with longer forecast periods. This LSTM model may be used with a weather parameter prediction model to manage short-term lightning risk.