时空聚焦感知与协同加权优化的多任务闪电临近预报方法

唐志豪; 韩袁鹏; 张慧; 宋琳; 张其林; 刘毅

doi:10.11999/JEIT260234

时空聚焦感知与协同加权优化的多任务闪电临近预报方法

doi: 10.11999/JEIT260234 cstr: 32379.14.JEIT260234

唐志豪¹,
韩袁鹏¹,
张慧^{1, 2},
宋琳³,
张其林²,
刘毅^{1, 2}

1.
南京信息工程大学江苏省大气环境与装备技术协同创新中心南京 210044
2.
南京信息工程大学气候系统预测与变化应对全国重点实验室/气象灾害教育部重点实验室/气象灾害预报预警与评估协同创新中心南京 210044
3.
青岛气象局青岛市生态与农业气象中心(青岛市气候变化中心) 青岛 266003

详细信息

作者简介:
唐志豪：男，硕士生，研究方向为深度学习、计算机视觉

韩袁鹏：男，硕士生，研究方向为深度学习、计算机视觉

张慧：女，硕士生，研究方向为深度学习、计算机视觉

宋琳：女，高级工程师，研究方向为雷电预警、大气物理

张其林：男，教授、博士生导师，研究方向为雷电预警、气象探测仪器

刘毅：男，讲师，硕士生导师，研究方向为气象大数据、深度学习、计算机视觉

中图分类号: TP181
计量
- 文章访问数: 6
- HTML全文浏览量: 4
- PDF下载量: 0
- 被引次数: 0
出版历程
- 修回日期: 2026-06-15
- 录用日期: 2026-06-15
- 网络出版日期: 2026-06-19

Multi-Task Lightning Nowcasting with Spatio-Temporal Focal Perception and Synergistic Weighted Loss

1.
Collaborative Innovation Center on Atmospheric Environment and Equipment Technology, B-DAT, Nanjing University of Information Science and Technology, Nanjing 210044, China
2.
State Key Laboratory of Climate System Prediction and Risk Management/ Key Laboratory of Meteorological Disaster, Ministry of Education/ Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science and Technology, Nanjing 210044, China
3.
Qingdao Ecological and Agricultural Meteorological Center, Qingdao Meteorological Bureau, Qingdao 266003, China

摘要

摘要: 闪电临近预报对气象预警与设施安全至关重要。然而，现有深度学习方法常受限于数据的极端稀疏性，且预测目标多为二值落区而非频次与区域的协同优化。文章提出一种多任务临近预报模型——时空聚焦感知与协同加权损失网络(STF-Net)，实现闪电频次与区域的联合预测。首先，设计闪电自适应注意力模块(LAAM)，显式建模长程时空依赖并精准聚焦对流敏感区；其次，构建时空加权混合损失函数，联合优化时间加权均方误差与双重加权交叉熵(DWCE)，以有效抑制极端稀疏分布诱发的优化偏误与虚警；最后，引入时空双分支生成对抗机制，提升预测场的细节保真度与时序连贯性。基于高分辨率甚低频闪电数据集的实验表明，STF-Net在1小时预报窗口内的临界成功指数(CSI)达0.663，较基线提升14.5%，虚警率(FAR)显著降至0.216，有效缓解了长时效预报的性能衰减。本研究为业务化闪电预警系统提供了一种高效、具备稀疏自适应能力的端到端解决方案。
- 闪电位置预报 /
- 闪电频次预报 /
- 时空加权损失函数 /
- 注意力机制 /
- 灾害预警
Abstract: Objective Lightning nowcasting is vital for early warning systems and the protection of critical infrastructure such as aviation, power grids, and transportation. Traditional numerical weather prediction models suffer from parameterization dependencies and high computational costs, making them inefficient for rapid-update nowcasting. Existing deep learning methods, despite progress, inadequately handle extreme data sparsity, suffer from serial computation bottlenecks in recurrent architectures, and primarily focus on binary occurrence prediction rather than the synergistic optimization of frequency estimation and regional localization. Moreover, conventional loss functions are easily dominated by vast non-lightning areas, causing biased predictions toward zero or generating excessive false alarms. To address these limitations, this paper proposes STF-Net, a multi-task lightning nowcasting model that jointly predicts lightning frequency and occurrence regions through three key innovations: a Lightning Adaptive Attention Module (LAAM) for explicit spatio-temporal dependency modeling, a Spatio-Temporally Weighted Hybrid Loss function to tackle data sparsity and imbalance, and a spatio-temporal dual-branch Generative Adversarial Network (GAN) to enhance prediction fidelity and temporal coherence. Methods STF-Net is built upon the SimVP video prediction architecture, adopting an encoder-translator-decoder paradigm. The Lightning Adaptive Attention Module (LAAM) employs a three-dimensional decoupled attention mechanism along height, width, and channel dimensions, enabling adaptive focus on convectively sensitive regions while maintaining computational efficiency. The Spatio-Temporally Weighted Hybrid Loss combines Temporally-Weighted Mean Squared Error (TW-MSE) for frequency regression accuracy and Dual-Weighted Cross-Entropy Loss (DWCE) for precise regional identification, incorporating time-increasing weights to enhance medium-to-long-term forecast robustness (Fig. 5). The DWCE loss innovatively integrates static class weights with dynamic grid weights, effectively balancing global class proportions and local lightning frequency heterogeneity.A spatio-temporal dual-branch GAN, comprising a spatial PatchGAN discriminator and a temporal 3D convolutional discriminator, is introduced to improve the textural fidelity and temporal coherence of the predicted lightning frequency fields. The model processes 6 consecutive historical lightning frequency frames (256×256 resolution, 10-minute intervals) to predict the next 6 frames, corresponding to a 1-hour forecast window. Experiments are conducted on a high-resolution Very Low Frequency Lightning Location Network (VLF-LLN) dataset containing 11,748 images covering diverse seasonal and weather conditions, split 7:3 for training and testing. Results and Discussions Comprehensive evaluation metrics are employed, including frequency regression accuracy (MSE, MAE computed only on lightning-occurring pixels), image quality fidelity (PSNR, SSIM), and regional detection skills (POD, FAR, CSI). STF-Net achieves a Critical Success Index (CSI) of 0.663 within the 1-hour forecast window, a 14.5% improvement over the SimVP baseline (0.579), and reduces the False Alarm Rate (FAR) from 0.351 to 0.216, a relative reduction of 38.5% (Table 1). Ablation studies systematically validate each component: adding GAN improves CSI to 0.624 and reduces MSE to 0.109; further incorporating LAAM increases CSI to 0.629 with the highest POD of 0.894; the complete STF-Net with Hybrid Loss achieves optimal performance with CSI of 0.663 and MSE of 0.105 (Table 1, Fig. 5). Critically, the synergistic prediction of frequency and region is evidenced by the simultaneous improvement in both regression (MSE/MAE) and detection (CSI/FAR) metrics. Time-step analysis reveals that LAAM significantly mitigates long-term performance degradation, with STF-Net maintaining the highest CSI compared to SimVP+GAN and SimVP (Fig. 6). Comparative experiments against ConvLSTM and PredRNN demonstrate STF-Net's superiority across all lead times: it consistently achieves higher CSI and lower FAR, with the advantage becoming more pronounced as the forecast horizon extends (Fig. 7). The consistent advantage in PSNR and SSIM further underscores the spatio-temporal GAN's role in producing coherent, detail-rich predictions.Visualization results show that STF-Net generates structurally clear, continuous lightning activity regions centered on high-frequency areas, accurately tracking dynamic evolution patterns including movement, merging, and splitting while producing minimal noise in non-lightning regions, demonstrating effective collaborative prediction of both frequency magnitude and spatial distribution . Conclusions This paper presents STF-Net, a novel deep learning model that achieves synergistic lightning frequency and regional prediction through three core contributions: (1) the Lightning Adaptive Attention Module (LAAM) explicitly models long-range spatio-temporal dependencies and focuses on critical convective zones; (2) the Spatio-Temporally Weighted Hybrid Loss effectively addresses extreme data sparsity and class imbalance, simultaneously optimizing frequency regression accuracy and regional localization precision while suppressing false alarms; and (3) the spatio-temporal dual-branch GAN enhances the spatial structural consistency of predictions and temporal coherence of predictions. Experimental results demonstrate that STF-Net significantly outperforms baseline and state-of-the-art models, achieving superior CSI of 0.663, reduced FAR of 0.216, and optimal MSE/MAE values within a 1-hour forecast window. The model effectively mitigates long-term performance degradation, accurately captures lightning region evolution trends, and generates physically plausible predictions with minimal background noise. This research provides an efficient, end-to-end solution for operational lightning nowcasting systems and offers new insights for model design in sparse meteorological spatio-temporal sequence prediction.
- Lightning location forecasting /
- Lightning frequency forecasting /
- Spatio-temporal weighted loss function /
- Attention mechanism /
- Disaster early warning

HTML全文

图 1 原始闪电命中频次与高斯模糊处理后效果图

下载: 全尺寸图片幻灯片

图 2 STF-Net模型的总体架构

下载: 全尺寸图片幻灯片

图 3 闪电自适应注意力模块架构图

下载: 全尺寸图片幻灯片

图 4 深层时空特征提取模块结构图

下载: 全尺寸图片幻灯片

图 5 含有不同模块的模型闪电预报对比效果图

下载: 全尺寸图片幻灯片

图 6 不同模型变体的预报性能对比

下载: 全尺寸图片幻灯片

图 7 各模型预报的可视化结果

下载: 全尺寸图片幻灯片

图 8 各模型的预报性能对比

下载: 全尺寸图片幻灯片

表 1 消融实验结果

评价指标	Baseline (SimVP)	+GAN	+LAAM	Light weight	STF-Net
MSE ↓	0.134	0.109	0.109	0.112	0.105
MAE ↓	0.048	0.048	0.047	0.044	0.043
POD ↑	0.862	0.872	0.894	0.787	0.819
FAR ↓	0.351	0.329	0.330	0.230	0.216
CSI ↑	0.579	0.624	0.629	0.641	0.663

下载: 导出CSV

表 2 不同模型预报的各指标结果对比

评价指标	ConvLSTM	PredRNN	STF-Net
MSE ↓	0.131	0.115	0.105
MAE ↓	0.056	0.048	0.043
PSNR ↑	37.77	41.44	42.42
SSIM ↑	0.8903	0.9219	0.9452
CSI ↑	0.497	0.622	0.663

下载: 导出CSV

参考文献(24)

[1]	MISHRA M, ACHARYYA T, SANTOS C A G, et al. Mapping main risk areas of lightning fatalities between 2000 and 2020 over Odisha state (India): A diagnostic approach to reduce lightning fatalities using statistical and spatiotemporal analyses[J]. International Journal of Disaster Risk Reduction, 2022, 79: 103145. doi: 10.1016/j.ijdrr.2022.103145.
[2]	SONG Yang, XU Cangsu, LI Xiaolu, et al. Lightning-induced wildfires: An overview[J]. Fire, 2024, 7(3): 79. doi: 10.3390/fire7030079.
[3]	王昊亮, 刘玉宝, 赵天良, 等. 基于数值天气模式及其模式输出的闪电预报研究进展[J]. 地球科学进展, 2017, 32(1): 44–55. doi: 10.11867/j.issn.1001-8166.2017.01.0044. WANG Haoliang, LIU Yubao, ZHAO Tianliang, et al. Progress in lightning forecast by using numerical weather models and model outputs[J]. Advances in Earth Science, 2017, 32(1): 44–55. doi: 10.11867/j.issn.1001-8166.2017.01.0044.
[4]	PARAMANIK M M R, RABBANI K M G, IMRAN A, et al. Prediction of lightning activity over Bangladesh using diagnostic and explicit lightning parameterizations of WRF model[J]. Natural Hazards, 2024, 120(5): 4399–4422. doi: 10.1007/s11069-023-06355-6.
[5]	FEDERICO S, TORCASIO R C, LAGASIO M, et al. A year-long total lightning forecast over Italy with a dynamic lightning scheme and WRF[J]. Remote Sensing, 2022, 14(14): 3244. doi: 10.3390/rs14143244.
[6]	BATTAGLIOLI F, GROENEMEIJER P, TSONEVSKY I, et al. Forecasting large hail and lightning using additive logistic regression models and the ECMWF reforecasts[J]. Natural Hazards and Earth System Sciences, 2023, 23(12): 3651–3669. doi: 10.5194/nhess-23-3651-2023.
[7]	谈玲, 康瑞星, 夏景明, 等. 融合多源异构气象数据的光伏功率预测模型[J]. 电子与信息学报, 2024, 46(2): 503–517. doi: 10.11999/JEIT230731. TAN Ling, KANG Ruixing, XIA Jingming, et al. A photovoltaic power prediction model integrating multi-source heterogeneous meteorological data[J]. Journal of Electronics & Information Technology, 2024, 46(2): 503–517. doi: 10.11999/JEIT230731.
[8]	孙强, 赵珂. 嵌入多阶泰勒微分知识的多尺度注意力循环网络深度时空序列预测方法[J]. 电子与信息学报, 2024, 46(6): 2605–2618. doi: 10.11999/JEIT231108. SUN Qiang and ZHAO Ke. Multi-scale attention recurrent network with multi-order Taylor differential knowledge for deep spatiotemporal sequence prediction[J]. Journal of Electronics & Information Technology, 2024, 46(6): 2605–2618. doi: 10.11999/JEIT231108.
[9]	. SHI Xingjian, CHEN Zhourong, WANG Hao, et al. Convolutional LSTM network: A machine learning approach for precipitation nowcasting[C]. Proceedings of the 29th International Conference on Neural Information Processing Systems, Montreal, Canada, 2015: 802–810.
[10]	CINTINEO J L, PAVOLONIS M J, and SIEGLAFF J M. ProbSevere LightningCast: A deep-learning model for satellite-based lightning nowcasting[J]. Weather and Forecasting, 2022, 37(7): 1239–1257. doi: 10.1175/WAF-D-22-0019.1.
[11]	. WANG Yunbo, LONG Mingsheng, WANG Jianmin, et al. PredRNN: Recurrent neural networks for predictive learning using spatiotemporal LSTMs[C]. Proceedings of the 31st International Conference on Neural Information Processing Systems, Long Beach, USA, 2017: 879–888.
[12]	黄兴友, 马玉蓉, 胡苏蔓. 基于深度学习的天气雷达回波序列外推及效果分析[J]. 气象学报, 2021, 79(5): 817–827. doi: 10.11676/qxxb2021.041. HUANG Xingyou, MA Yurong, and HU Suman. Extrapolation and effect analysis of weather radar echo sequence based on deep learning[J]. Acta Meteorologica Sinica, 2021, 79(5): 817–827. doi: 10.11676/qxxb2021.041.
[13]	LI Fengquan, LI Jian, GU Shanqiang, et al. Nowcasting of cloud-to-ground lightning location and frequency based on a deep learning technique[J]. Atmospheric and Oceanic Science Letters, 2025, 19(3): 100607. doi: 10.1016/j.aosl.2025.100607.
[14]	任照环, 林锐, 周浩, 等. 基于深度学习模型的雷电落区预报[J]. 热带气象学报, 2024, 40(5): 758–766. doi: 10.16032/j.issn.1004-4965.2024.068. REN Zhaohuan, LIN Rui, ZHOU Hao, et al. Lightning strike location prediction using a deep learning model[J]. Journal of Tropical Meteorology, 2024, 40(5): 758–766. doi: 10.16032/j.issn.1004-4965.2024.068.
[15]	ARIFIN S, SAKIB N, RAYHAN Y, et al. Lightning prediction under uncertainty: Deeplight with hazy loss[J]. Expert Systems with Applications, 2026, 314: 131586. doi: 10.1016/j.eswa.2026.131586.
[16]	. KUMAR R, SHARMA T, VAGHELA V, et al. PrecipFormer: Efficient transformer for precipitation downscaling[C]. Proceedings of the 2025 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), Tucson, USA, 2025: 452–460. doi: 10.1109/WACVW65960.2025.00056.
[17]	职为梅, 常智, 卢俊华, 等. 面向不平衡图像数据的对抗自编码器过采样算法[J]. 电子与信息学报, 2024, 46(11): 4208–4218. doi: 10.11999/JEIT240330. ZHI Weimei, CHANG Zhi, LU Junhua, et al. Adversarial autoencoders oversampling algorithm for imbalanced image data[J]. Journal of Electronics & Information Technology, 2024, 46(11): 4208–4218. doi: 10.11999/JEIT240330.
[18]	CAVAIOLA M, CASSOLA F, SACCHETTI D, et al. Hybrid AI-enhanced lightning flash prediction in the medium-range forecast horizon[J]. Nature Communications, 2024, 15(1): 1188. doi: 10.1038/s41467-024-44697-2.
[19]	. GAO Zhangyang, TAN Cheng, WU Lirong, et al. SimVP: Simpler yet better video prediction[C]. Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, 2022: 3160–3170. doi: 10.1109/CVPR52688.2022.00317.
[20]	TIAN Jie, HAN Wei, SUN Haofei, et al. A dual-weighted loss function for lightning nowcasting[J]. Atmospheric and Oceanic Science Letters, 2026, 19(3): 100778. doi: 10.1016/j.aosl.2026.100778.
[21]	LI Jie, DAI Bingzhe, ZHOU Jiahao, et al. Preliminary application of long-range lightning location network with equivalent propagation velocity in China[J]. Remote Sensing, 2022, 14(3): 560. doi: 10.3390/rs14030560.
[22]	SRIVASTAVA A, LIU Dongxia, XU Chen, et al. Lightning nowcasting with an algorithm of thunderstorm tracking based on lightning location data over the Beijing area[J]. Advances in Atmospheric Sciences, 2022, 39(1): 178–188. doi: 10.1007/s00376-021-0398-2.
[23]	周康辉, 郑永光, 蓝渝. 基于闪电数据的雷暴识别、追踪与外推方法[J]. 应用气象学报, 2016, 27(2): 173–181. doi: 10.11898/1001-7313.20160205. ZHOU Kanghui, ZHENG Yongguang, and LAN Yu. Flash cell identification, tracking and nowcasting with lightning data[J]. Journal of Applied Meteorological Science, 2016, 27(2): 173–181. doi: 10.11898/1001-7313.20160205.
[24]	LIU Yi, WANG Jiale, SONG Yang, et al. Lightning nowcasting based on high-density area and extrapolation utilizing long-range lightning location data[J]. Atmospheric Research, 2025, 321: 108070. doi: 10.1016/j.atmosres.2025.108070.