LightNet+: A dual-source lightning forecasting network with bi-direction spatiotemporal transformation

Xinyuan Zhou, Yangli-ao Geng, Haomin Yu, Qingyong Li, Liangtao Xu, Wen Yao, Dong Zheng, Yijun Zhang

Abstract: Lightning disaster causes a huge threat to human lives and industrial facilities. Data-driven lightning forecasting plays an effective role in alleviating such disaster losses. The forecasting process usually faces multi-source meteorological data characterized by spatiotemporal structure. However, established data-driven forecasting methods are mostly built on classic convolutional and recurrent neural blocks which processes one local neighborhood at a time, failing to capture long-range spatiotemporal dependencies within data. To address this issue, we propose a dual-source lightning forecasting network with bi-direction spatiotemporal transformation, referred to as LightNet $$+$$ + . The core of LightNet $$+$$ + is a novel module, namely bi-directional spatiotemporal propagator, which aims to model long-range connections among different spatiotemporal locations, going beyond the constraints of the receptive field of a local neighborhood. Moreover, a spatiotemporal encoder is introduced to extract historical trend information from recent observation data. Finally, all the obtained features are organically fused via a non-local spatiotemporal decoder, which then produces final forecasting results. We evaluate LightNet $$+$$ + on a real-world lightning dataset from North China and compare it with several state-of-the-art data-driven lightning forecasting methods. Experimental results show that the proposed LightNet $$+$$ + yields overall best performance.