VLF/LF Lightning Signals Generation Based on CNN-Transformer Network

This paper proposes a CNN-Transformer hybrid network for generating VLF/LF lightning electromagnetic pulse signals. To address the limitations of traditional Finite-Difference Time-Domain (FDTD) methods, such as low computational efficiency and real-time simulation challenges, we establish a deep learning-based mapping between lightning parameters (peak current, distance, azimuth angle, time) and waveforms. Using lightning electromagnetic field waveform data collected by the 3D lightning detection system of the Institute of Electrical Engineering, Chinese Academy of Sciences, the model extracts features through 1D convolutional layers and global dependencies via multi-head self-attention mechanisms. Experimental results show that 81.8% of generated waveforms achieve a cosine similarity exceeding 0.6 with real data, 10.9% exceed 0.9, and 77.9% have peak time offsets within ±5 μs. The influence of input parameters on output waveforms aligns with empirical observations. This method provides a data simulation tool for validating lightning detection systems and advances deep learning applications in lightning electromagnetic pulse waveform processing.