A Graph Neural Network Based Workflow for Real-Time Lightning Location With Continuous Waveforms

Abstract

Real-time lightning monitoring is crucial for public safety and infrastructure protection by quickly and accurately locating lightning strikes. Traditional methods, such as time of arrival algorithms, rely on precise arrival time picking, which can compromise localization accuracy. Conversely, time reversal (TR) algorithms bypass picking but are hindered by time-consuming grid search requirements. We propose a graph neural network (GNN) based method for accurate, real-time multi-station lightning location. Our approach processes continuous lightning waveform data from multiple sensors, achieving simultaneous denoising, event detection, and direct localization without arrival time picking. Specifically, denoising enhances the signal-to-noise ratio, thereby improving the accuracy of subsequent event detection and localization. Events are identified by matching signals across sensors and retaining high-match segments, resulting in waveforms that contain valid lightning signals. These waveforms are then input into a GNN, which integrates time series features with spatial information from the sensors, effectively handling multi-station localization and delivering accurate, real-time results. To address the lack of training data sets for lightning location, we propose a novel procedure for constructing a labeled lightning data set, laying a data foundation for future data-driven approaches in this domain. In extensive synthetic experiments, our method achieved a low average localization error of 0.61 km and high efficiency with a localization time of only 0.4 milliseconds, significantly outperforming the traditional TR algorithm's 1.16 km error and 1.65 s. When tested on natural cloud-to-ground lightning data, our method successfully detected and located 198 lightning sources consistent with reference results.