Short range lightning forecast based on deep learning models using globally available gridded meteorological data

Abstract

In the context of climate change and its effects on extreme weather events, the prediction and monitoring of lightning strikes are crucial, especially in regions prone to such events. In this context, this study aims to develop and evaluate deep learning models to predict lightning strikes up to 6 hours in advance, using globally available meteorological data from the Global Data Assimilation System (GDAS) as input, and lightning occurrence data provided by the Atmospheric Electricity Group (ELAT) of the National Institute for Space Research (INPE) as the target. The state of Bahia, Brazil, was selected as a case study due to its vast region where many electrical projects are being developed, justifying the study and application of this research to that region. Several deep learning models were tested to identify an accurate and high-performing model for early prediction of light ning occurrences. In total, six deep learning models were considered, including Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and hybrid networks. These models were optimized to achieve maximum predic tion accuracy. Each model was meticulously trained and evaluated, and its performance was assessed using traditional classification metrics, including F1-score, Area Under the Receiver Operating Characteristic Curve (ROC AUC), Area Under the Precision-Recall Curve (PRC AUC), precision, recall, specificity, and accuracy. The Delong test was used to evaluate statistical significance and detect differences between the models. The study successfully identified models that outperformed others in predicting both lightning and no lightning cases, and these models were found to be significantly distinct from each other. Specifically, the CNN-GRU model demonstrated superiority compared to the oth ers, with good computational efficiency. Additionally, potential limitations were identified, resulting in recommendations for future work in lightning prediction using deep learning. Based on the results, it can be concluded that advanced deep learning techniques can significantly contribute to accurate lightning occurrence predictions with good compu tational performance. By identifying the most suitable models for lightning prediction, this research serves as a foundation for the development of more effective alert systems, contributing to the safety of society and industry, as well as climate resilience.