Thunderstorm severity prediction for South Africa between 2015 to 2016 using LSTM model variants and remote sensing instruments
Lightning is one of the leading causes of electrical outages in South Africa, and the most severe weather-related killer in the country. Unfortunately, quantitative lightning prediction remains challenging. In this research, we evaluate the accuracy of LSTM neural network model variants on lightning frequency prediction using remote sensing instruments. These LSTM model variants are the LSTM-Fully Connected (LSTM-FC), the Convolutional-Neural-Network-LSTM (CNN-LSTM) and the Convolutional-LSTM (ConvLSTM) models. Both the CNN-LSTM model and the ConvLSTM model recognize spatiotemporal features. The data consists of lightning detection data from the South African Lightning Detection Network (SALDN) and weather-feature information from the South African Weather Service. We forecast lighting frequency, every hour, between December-2013 and March-2016 within the predicted area. The models used in this research were trained on data between July-2008 to November 2013. All models were trained to minimize Mean Squared Error but were evaluated on Mean Absolute Error (MAE flashes.hr-1). We also varied models based on input datasets: SALDN-only, SAWS-only and SALDN+SAWS datasets. We found that the CNN-LSTM model (MAE=51) performed best amongst LSTM model variants (LSTM-FC MAE=67; ConvLSTM MAE=86). When models were evaluated between input datasets, we found that SALDN only (MAE= 59) outperformed SAWS only and SALDN+SAWS (SAWS MAE=74; SAWS+SALDN MAE=70). We conclude that CNN-LSTM models outperform prediction accuracy for thunderstorm severity compared with ConvLSTM and FC-LSTM models but consideration on input data is required.
A dissertation in partial fulfilment for the degree of Master of Science in the field of e-Science, Faculty of Science, School of Computer Science and Applied Mathematics, University of the Witwatersrand, Johannesburg, 2022