Abstract Title: A Novel NWP-AI Hybrid Approach For Lightning Early Warning System

Abstract Submitted to: ATMOSPHERIC AND SPACE ELECTRICITY

Abstract Text:

The atmospheric processes associated with lightning are yet to be fully understood, making it challenging to forecast using numerical weather prediction (NWP) models with significant skill. As artificial intelligence (AI) is becoming one of the most common technologies in recent years, scientists focus on incorporating AI for weather forecasting. However, the highly imbalanced nature of the lightning-to-non-lightning ratio poses one major constraint for AI-based models. Gridded lightning flash density derived from the Indian Institute of Tropical Meteorology (IITM) Lightning Location Network (LLN) shows that between 2019 and 2020, over Maharashtra, India, lightning was only 0.20 % of the total dataset. To address the “curse of imbalance dataset”, we have developed a two-autoencoder-based classifier model (C2AE). C2AE, a deep-learning-based algorithm, is trained to correctly encode meteorological features associated with lightning and non-lightning incidents. For examining our methodology, the Indian Meteorology Department (IMD) WRF operational forecasts at 9 km resolution with 00 UTC initial condition and 24-hour lead time has been selected as input for the deep learning model. IITM LLN data is the target vector for training and verification of C2AE. When tested on a new dataset consisting of lightning and non-lightning data, C2AE has been able to predict lightning with statistically significant skills. Hence this study shows that a simple deep-learning algorithm can forecast lightning with sufficient skill using WRF 24-hour lead time at 9 km resolution, a considerable step toward building lightning early warning systems in India.