Application of the VLF-LLN Data and Lightning Nowcasting Based on Multi-Source Fusion Model

Accurate and timely lightning nowcasting is critical for industry safety and public safety.  Based on the data from the self-built Very Low Frequency Long-Range Lightning Location Network (VLF-LLN), combined with radar data and Himawari-8 satellite data, our team researched the lightning nowcasting within 0-1 hours in China using a number of deep learning models.

• A new algorithm for the long-range lightning location.The algorithm is used to extract the ground wave peak points of the lightning sferics based on the waveform bank and extracted the arrival time of the ground wave, and the lighting flashes can been located by using TOA (Time of Arrival) combined with the equivalent propagation velocity method, and the detection efficiency of the VLF-LLN was determined to be up to 70.6% with a median location accuracy of 1.814 km.
• Lightning nowcasting based on high-density area and extrapolation only utilizing VLF-LLN data.More current lightning nowcasting is based on data-driven methods, which mainly focus on the optimization of model structure, ignoring the characteristic factors of the data itself, especially the discontinuity in time and the discreteness in space of lightning location data. Therefore, a Gaussian kernel with a radius of 20 km is first used to diffuse the original lightning strike frequency image, and through the weight allocation between adjacent frames, we obtained a sequence of mutually related lightning image frames, and separately used the frame data for lightning recognition and extrapolation. The results show that the accuracy rates of lightning areas extrapolation within the next 6 minutes, 18 minutes, and 36 minutes are 94.1%, 79.4%, and 65.8%, respectively.
• Lightning nowcasting based on Himawari-8 satellite data, radar data, and VLF-LLN data.Based on the deep learning models, we have solved the problems of predicting the first lightning and extrapolation for the lightning nowcasting. We consider time as an additional channel dimension and combine it with recent advances in attention mechanisms, a lightweight model was proposed to obtain better performance index than other baseline models with minimal computational resource, which consists of an encoder and decoder based on two-dimensional convolution and several temporal translators based on one-dimensional convolution. The data include water vapour in the middle troposphere and cloud top temperature from the Himawari-8 satellite, radar combined reflectivity factor, and VLF-LLN lightning data. The results indicate that the radar data significantly improves the hit rate and reduces the false alarm rate, while satellite data mainly reduces the false alarm rate. The average hit rate, false alarm rate and critical success index for lightning nowcasting are 0.5645, 0.3302 and 0.4441, respectively.