image: The ERA5 data for the past hour is first input into the weather foundation model to generate forecasts for the next Tf hours. These forecasts are then cropped to focus on a specific local region. In addition, recent lightning observation data over the past Tp hours for the forecasting region is incorporated. Finally, the gSTFNet is used to integrate these inputs and generate lightning forecasts for the next Tf hours. The gSTFNet consists of four parts: the WFM data encoder (Module I), the observation encoder (Module II), the gated spatiotemporal fusion module (Module III), and the forecasting decoder (Module IV).
Credit: ©Science China Press
This study is led by Li Yiran, Dr. Li Qingyong, Dr. Geng Yangli-ao and Guo zhiqing from Beijing Jiaotong University, and Dr. Zheng Dong, Dr. Xu Liangtao, Yao wen and Dr. Lyu Weitao from Chinese Academy of Meteorological Sciences. The researchers studies a refined lightning forecasting method based on WFMs (specifically adopting Pangu-Weather) and proposes a dual-source forecasting framework that combines the complementary strengths of WFMs in long-term weather trend forecasting and recent lightning observations in short-term extrapolation to enhance forecasting performance. Methodologically, the team designs a gated spatiotemporal fusion network (gSTFNet) to address the challenges of cross-temporal and cross-modal fusion of dual-source data. The dual-source encoder is first used to separately encode features from WFMs and lightning observations, effectively narrowing the modal gap in the latent feature space. The gated spatiotemporal fusion module then integrates the dual-source features in the latent space, facilitating seamless cross-temporal fusion. The forecasting decoder is finally employed to enhance the spatial resolution of the fused features, generating high-resolution lightning forecasts.
In evaluation on real-world lightning observation data collected in Guangdong from 2018 to 2022, gSTFNet stands out with significantly better forecast accuracy than the ECMWF HRES product and six deep-learning spatiotemporal forecasting baselines. This advantage can be attributed to the gated spatiotemporal fusion module in gSTFNet, which more effectively captures the spatiotemporal dependencies of dual-source features, and fully integrates the complementary strengths of dual-source data in long-term weather trend forecasting and short-term extrapolation, thereby enhancing the overall forecasting capability of the model.
To examine the impact of different data sources on the forecast performance of the gSTFNet model, the research team compared the evaluation results of gSTFNet-P (an ablation version of gSTFNet trained using only WFMs forecasts) and gSTFNet-L (an ablation version of gSTFNet trained using only lightning observations). The experimental results clearly demonstrate that gSTFNet-P outperforms HRES, suggesting that the forecast accuracy of WFMs is comparable to that of the current leading NWP method. In contrast, gSTFNet-L significantly outperforms gSTFNet-P in short term forecasting, due to the strong correlation between past and future lightning occurrences, which the model effectively captures and models through training. However, as the forecast duration increases, gSTFNet-P’s accuracy becomes more stable, whereas gSTFNet-L’s performance declines rapidly and ultimately falls below that of gSTFNet-P. By combining the strengths of the WFMs in forecasting long-term weather trends with the supplementary role of lightning observations in short-term forecasting, gSTFNet outperforms the other models across nearly all time intervals, further validating the effectiveness of this integration approach.
In summary, although current WFMs do not directly generate lightning forecasts, adapting their outputs through neural network training has yielded promising results. However, relying solely on WFMs outputs does not significantly improve lightning forecasting accuracy. A more effective approach is to combine WFMs forecasts with recent lightning observations, thereby leveraging the complementary strengths of both data sources. This integration substantially enhances forecast performance.
See the article:
Li Y, Li Q, Zheng D, Geng Y, Guo Z, Xu L, Yao W, Lyu W. 2025. A gated spatiotemporal fusion network for lightning forecasting based on weather foundation models. Science China Earth Sciences, 68(9): 2957–2975, https://doi.org/10.1007/s11430-025-1638-8
Journal
Science China Earth Sciences