III: Small: Prediction and Characterization of Extreme Events in Spatio-Temporal Data.

III：小：时空数据中极端事件的预测和表征。

• 批准号: 2006633
• 负责人: Pang-Ning Tan
• 金额: $ 49.99万
• 依托单位: Michigan State University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-10-01 至 2024-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2006633&HistoricalAwards=false
• 关键词: III; Small; Prediction; Characterization; Extreme

Extreme weather and climate events such as hurricanes, heat waves, and droughts are destructive natural forces with the potential to cause devastating losses in property and human lives. According to the National Center for Environmental Information (NCEI), there have been more than 40 weather and climate disasters in the United States since 2017 that cost over $1 billion each, incurring over $460 billion in total losses and more than 3500 deaths. Given the severity of their impact, accurate prediction of the magnitude, frequency, timing, and location of such extreme events are critical to provide timely information to the public and to minimize the risk for human casualties and property destruction, thus advancing the national health, prosperity and welfare. However, despite their importance, forecasting the extreme events from spatio-temporal data is still a great challenge as the events to be detected are often rare and hard to predict. Identifying the spatio-temporal drivers of the extreme events is also a challenge as the events typically involve complex, nonlinear interactions between the underlying natural and anthropogenic processes. Through development and use of machine learning algorithms, this project will contribute to the advances of science to better predict these extreme events. This project aims to develop novel algorithms for predicting and characterizing extreme events in large-scale spatio-temporal data. Specifically, the planned research combines statistical theories for extreme value distribution with deep learning to enable accurate prediction and characterization of the extreme events. To achieve this goal, the planned research centers around the following three key areas: (1) development of deep learning algorithms with extreme value theory for predicting and characterizing extreme events in time series forecasting problems, (2) development of convolutional methods for joint extreme event forecasting at multiple locations, and (3) development of extreme event prediction methods for spatial trajectory data. As proof of concept, the planned methods will be applied to a variety of environmental monitoring applications, including the prediction of extreme weather events such as heat waves, droughts, and hurricanes. The planned research is transformative as it will shed light on the following key issues: (1) How to bridge the gap between current extreme value theory for modeling the tail distribution of random phenomena with deep learning. (2) How to design spatio-temporal deep learning approaches that can accurately forecast the magnitude, frequency, and timing of extreme events at multiple locations. and (3) How to design a deep learning framework for extreme value prediction in spatial trajectory data. Successful completion of this project will be a significant step forward towards resolving these issues.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

飓风、热浪和干旱等极端天气和气候事件是破坏性的自然力量，有可能造成毁灭性的财产和生命损失。根据国家环境信息中心(NCEI)的数据，自2017年以来，美国已经发生了40多起天气和气候灾害，每个灾害的损失超过10亿美元，总损失超过4600亿美元，超过3500人死亡。鉴于其影响的严重性，准确预测此类极端事件的规模、频率、时间和地点，对于及时向公众提供信息，将人员伤亡和财产破坏的风险降至最低，从而促进国家的健康、繁荣和福祉至关重要。然而，尽管极端事件很重要，但从时空数据中预测极端事件仍然是一个巨大的挑战，因为要检测的事件往往很少，而且很难预测。确定极端事件的时空驱动因素也是一项挑战，因为这些事件通常涉及潜在的自然和人为过程之间复杂的、非线性的相互作用。通过开发和使用机器学习算法，该项目将为更好地预测这些极端事件的科学进步做出贡献。该项目旨在开发新的算法来预测和描述大规模时空数据中的极端事件。具体地说，计划中的研究将极值分布的统计理论与深度学习相结合，以实现对极端事件的准确预测和表征。为了实现这一目标，计划的研究集中在以下三个关键领域：(1)利用极值理论开发用于预测和表征时间序列预测问题中的极端事件的深度学习算法，(2)开发用于多位置联合极端事件预测的卷积方法，以及(3)开发针对空间轨迹数据的极端事件预测方法。作为概念的证明，计划中的方法将应用于各种环境监测应用，包括对热浪、干旱和飓风等极端天气事件的预测。这项计划中的研究具有变革性，因为它将揭示以下关键问题：(1)如何弥合当前极值理论与深度学习在随机现象尾部分布建模方面的差距。(2)如何设计能够准确预测多个地点极端事件的强度、频率和时间的时空深度学习方法。(3)如何设计空间轨迹数据极值预测的深度学习框架。这个项目的成功完成将是朝着解决这些问题迈出的重要一步。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

DeepGPD: A Deep Learning Approach for Modeling Geospatio-Temporal Extreme Events

• DOI: 10.1609/aaai.v36i4.20344
• 发表时间: 2022-06
• 作者: T. Wilson;Pang-Ning Tan;L. Luo

COMET Flows: Towards Generative Modeling of Multivariate Extremes and Tail Dependence

COMET 流：走向多元极值和尾部依赖性的生成模型

• DOI: 10.24963/ijcai.2022/462
• 发表时间: 2022
• 期刊: Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence
• 作者: McDonald, Andrew;Tan, Pang-Ning;Luo, Lifeng
• 通讯作者: Luo, Lifeng

Self-Recover: Forecasting Block Maxima in Time Series from Predictors with Disparate Temporal Coverage Using Self-Supervised Learning

• DOI: 10.24963/ijcai.2023/414
• 发表时间: 2023-08
• 作者: A. Galib;Andrew McDonald;Pang-Ning Tan;L. Luo

Beyond Point Prediction: Capturing Zero-Inflated & Heavy-Tailed Spatiotemporal Data with Deep Extreme Mixture Models

• DOI: 10.1145/3534678.3539464
• 发表时间: 2022-08
• 期刊: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining
• 作者: T. Wilson;Andrew McDonald;A. Galib;Pang-Ning Tan;L. Luo

JOHAN: A Joint Online Hurricane Trajectory and Intensity Forecasting Framework

JOHAN：联合在线飓风轨迹和强度预报框架

• DOI: 10.1145/3447548.3467400
• 发表时间: 2021
• 期刊: KDD '21: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining
• 作者: Wang, Ding;Tan, Pang-Ning
• 通讯作者: Tan, Pang-Ning