In this study, we present LightningNet, a cutting-edge deep learning model tailored for precise lightning nowcasting in the vicinity of Shanghai. With its ability to forecast up to 96 min into the future, LightningNet employs a finely tuned 1 km resolution grid and a 12-min temporal resolution. The model’s core strength lies in its transformer architecture, which allows for seamless integration and processing of multi-source observation data, including Doppler weather radar network feeds, lightning location system, and atmospheric variables (temperature, horizontal winds, specific humidity and geopotential height). To further optimize its predictive capabilities, LightningNet undergoes pre-training, focused on two self-supervised objectives: extrapolation and mask-reconstruction. During this pre-training phase, we harness satellite datasets sourced from the Weather4cast competition, ensuring that our model is robustly trained to handle the nuances of real-world weather patterns. Overall, our successful implementation of transfer learning within this multi-modal model offers promising insights for advancing the field of lightning nowcasting.