《自然-通讯》｜全球热带气旋预报的基准数据集与深度学习方法

atmos85

转自：“气象学家”公众号
https://mp.weixin.qq.com/s/M77ajtNiIvvgJjiDFqmN5Q

全球热带气旋预报的基准数据集与深度学习方法

Benchmark dataset and deep learning method for global tropical cyclone forecasting

Cheng Huang, Pan Mu, Jinglin Zhang, Sixian Chan, Shiqi Zhang, Hanting Yan, Shengyong Chen & Cong Bai

Huang, C., Mu, P., Zhang, J. et al. Benchmark dataset and deep learning method for global tropical cyclone forecasting. Nat Commun 16, 5923 (2025). https://doi.org/10.1038/s41467-025-61087-4

The development of a tropical cyclone is influenced by various environmental factors, such as the subtropical high and atmospheric circulation. These factors can all affect the future evolution and track of the tropical cyclone. Best viewed in color.

研究背景

热带气旋（TC）是强烈且动态的大气系统，通常起源于热带至温带纬度，对全球热量和动量分布起着关键调节作用。强大的热带气旋对海上船只和海上平台构成潜在灾难性影响，登陆时还会引发一系列自然灾害，如强风、风暴潮和洪水，造成重大经济损失和人员伤亡威胁。随着气候变化，过去几十年热带气旋的强度和破坏性呈上升趋势。为减轻这些灾害，准确预测热带气旋的路径和强度至关重要。然而，热带气旋预报仍然是一个重大科学挑战，其困难源于众多影响因素，如大尺度大气环流、副热带高压位置、TC内部压力结构和中心、海表面温度（SST）及其他环境条件，这些因素相互作用复杂，对TC演变的综合影响尚未完全理解。

It includes TropiCycloneNet Dataset (TCND) and TropiCycloneNet Model (TCNM). TCND includes various data with time series. For TCNM, the golden branch is the environment data (Env-Data) encoder, called Environment-Time Network (Env-T-Net). The blue branch is the inherent attributes data of tropical cyclone (Data1d) encoder, called 1D-Data Encoder. The red branch is the meteorological grid data (Data3d) encoder, called 3D-Data Encoder. The Roulette selects different generators by the probability array, the circular ring below GC-Net, from Generator Chooser Net. All predictions from selected generators constitute the multiple potential tendencies of tropical cyclone. Best viewed in color.

研究意义

本研究提出了一种新的深度学习方法和一个开放的多模态热带气旋数据集，旨在提高热带气旋路径和强度预测的准确性。通过整合气象知识和深度学习技术，该研究不仅为热带气旋预报领域带来了新的视角，还为吸引更多的研究人员参与该领域提供了资源和方法支持，有望加速数据驱动的热带气旋预测研究的发展。

The distribution of tropical cyclones (TCs) with varying intensities in TropiCycloneNet Dataset (TCND) across different sea areas, categorized as tropical depression (TD, 10.8–17.1 m/s), tropical storm (TS, 17.2–24.4 m/s), severe tropical storm (STS, 24.5–32.6 m/s), typhoon (TY, 32.7–41.4 m/s), strong typhoon (STY, 41.5–50.9 m/s), and super typhoon (SuperTY,  >51.0 m/s). Sub-figures a–f illustrate the count of TCs with different intensities in specific areas. For example, in sub-figure a, the number 378 indicates there are 378 severe tropical storm intensity TCs in the Western North Pacific region in TCND. The number 1781 represents the total TCs of the Western North Pacific in TCND. Sub-figure g presents the count of TCs of various intensities across these six sea areas, while sub-figure h displays the total count of TCs in each sea area. Best viewed in color.

方法

研究中提出的TropiCycloneNet（TCN）包括两个主要部分：TropiCycloneNet数据集（TCND）和TropiCycloneNet模型（TCNM）。TCND是一个涵盖六个主要海洋盆地、包含70年多源数据的开放多模态TC数据集。TCNM是一个集成AI和气象学的预测模型，包含多个模块，如生成器选择网络和环境时间网络。该模型通过协同优化气象信息架构和数据集的全面时空覆盖，实现了对现有深度学习方法和官方气象预报的超越。

数据

TCND数据集是本研究的核心贡献之一。它不仅包括TC的固有属性数据（如经度、纬度、压力和风速），还包含了气象网格数据（如卫星图像和再分析数据）以及大量的TC环境数据。这些数据来自多个开放数据源，包括中国气象局热带气旋最佳路径数据集（CMA-BST）、国际最佳路径档案气候管理（IBTrACS）和欧洲中期天气预报中心（ECMWF）的ERA5再分析数据集。数据集的详细信息和处理方法在文中进行了详细描述，确保了数据的质量和适用性。

研究结果

通过在TCND数据集上的广泛实验，TCNM模型在多个指标上均优于现有的深度学习方法和官方气象预报。具体来说，TCNM在热带气旋路径和强度预测方面表现出色，尤其是在预测多潜在趋势方面，能够为气象学家和决策者提供更全面的预测信息。此外，TCNM在处理不同海洋区域的热带气旋时也显示出良好的适应性，这得益于其在训练过程中使用了全球范围内的多模态数据。

结论与不足

研究得出的主要结论是，通过整合多模态数据和气象知识，TCNM模型能够提供更准确的热带气旋路径和强度预测。然而，该研究也存在一些不足之处。例如，TCNM在长期路径预测方面表现不如短期预测，这可能与模型未能充分考虑TC的物理机制以及未能充分利用TCND中的所有信息有关。此外，尽管TCND是一个高质量的数据集，但在某些方面仍有改进空间，例如在南半球的数据稀缺问题和对复杂海洋环境（如南太平洋区域）的处理。

讨论

文章讨论了TCND和TCNM在热带气旋预测领域的潜在影响。TCND作为一个开放的多模态数据集，为AI研究人员提供了一个宝贵的资源，降低了进入该领域的门槛。TCNM模型则展示了深度学习在热带气旋预测中的潜力，尤其是在整合气象学和AI知识方面。此外，文章还讨论了如何利用大型气象模型（如PanguWeather和Fengwu）来进一步改进TCNM，以及如何通过结合不同的方法来提高预测的准确性。

atmos85

（接上篇）

未来工作

未来的研究方向可能包括探索如何利用深度学习技术模拟热带气旋的关键物理机制，以及如何进一步改进TCND数据集以更好地满足研究人员的需求。此外，结合TCNM与其他擅长预测热带气旋路径的方法（如PanguWeather模型和官方气象机构的其他方法）可能是未来的一个研究方向，这将有助于开发出能够更有效地利用TCND数据的更优模型，从而提高路径和强度预测的准确性。

The sea areas include Western North Pacific (WP), Eastern North Pacific (EP), North Atlantic (NA), North Indian (NI), South Indian (SI), and South Pacific (SP). ALL means the average errors of all the six sea areas. Sub-figure (a-1, 2, 3, 4) show the performance of our model on tropical cyclone (TC) track prediction. Sub-figure (b-1, 2, 3, 4) and (c-1, 2, 3, 4) show the performance of our model on TC intensity prediction. The TropiCycloneNet Model (TCNM) means that our model is trained on all the six sea areas. The means that only the data of WP are used for training our model. Metrics Distance, Pressure, and Wind are introduced in the section titled ‘Metrics’. Best view in color. Source data are provided as a Source Data file.

The comparative results are based on Western North Pacific (WP) sea area data, as most of the comparative methods only provide performance results for the WP region. Sub-figures (a-1, a-2, a-3) present the average absolute error comparisons of tropical cyclone (TC) predictions among different deep learning methods. Sub-figures (b-1, b-2, b-3) show the long-term forecasting (6-72 h) capabilities of TCNM. Sub-figures (c-1, c-2, c-3) demonstrate the performance of TCNM during rapid changes in TC track or intensity compared to its overall performance on the entire dataset. In these sub-figures, The three solid lines in the represent the results of Chinese Central Meteorological Observatory (CMO), Pangu-Weather, and TCNM on the complete test set, respectively. CMO-rInt, Pangu-rInt, and TCNM-rInt represent the results of CMO, Pangu-Weather, and TCNM on rapid intensifying TC cases, respectively. CMO-rTra, Pangu-rTra, and TCNM-rTra represent the results of CMO, Pangu-Weather, and TCNM on curving TC cases, respectively. Best viewed in color. Source data are provided as a Source Data file.

Examples of track and intensity predictions from 6 h to 24 h and the comparison between our method and previous deep learning model (MMSTN) on the potential predictions for six tropical cyclones: MALIKSI (2018, severe tropical storm, Summer), TRAMI (2018, super typhoon, Summer), PABUK (2019, severe tropical storm, Winter), FENGSHEN (2019, strong typhoon, Summer), ATSANI (2020, typhoon, Autumn), and MAYSAK (2020, super typhoon, Summer). Sub-figure a shows the track prediction performance of these two models and Sub-figure b shows the intensity prediction performance of these two models. Best viewed in color.

以上是对文章的详细解读，如有不当之处欢迎批评指出。