CASIA OpenIR  > 毕业生  > 博士学位论文
基于深度学习的短临降水预测方法研究
靳淇兆
2024-05-14
Pages108
Subtype博士
Abstract

降水是一种与人类生产生活密切相关的天气现象。准确预测短临降水不仅为农业管理、交通规划以及灾害预防等公共服务提供关键技术支持,也是一项具有挑战性的学术研究任务。近年来,深度学习在气象预测领域取得了重大突破。然而,现有的基于深度学习的短临降水预测方法主要利用雷达回波数据进行预测,忽略了降水过程中各种气象要素的影响。本文以多模态三维(高度、经度及纬度)气象数据为研究对象,探讨了基于深度学习的短临降水预测方法。这项研究不仅为降水预测提供了新的建模方式和技术手段,也探索了复杂时变条件下多模态时空数据的建模和非线性预测问题,具有重要的理论研究价值和广阔的应用前景。

由于气象系统的复杂性,构建基于多模态三维气象数据的深度网络进行短临降水预测面临多个挑战。首先,气象数据是一种复杂的多模态时空数据,其在多模态维度、空间维度和时间维度上普遍存在复杂的耦合关系。因此,如何建模气象数据的局部相关性是本文的首要任务。其次,气象数据在不同地区和季节会呈现出规律性变化,这种差异在时空位置相距较大时更为显著。因此,如何建模气象数据的时空特异性是本文需要解决的难点。此外,降水数据服从极端不平衡的长尾分布,这增加了模型预测极端降水的难度。如何在长尾分布下提升尾部类别的预测能力,是本文面临的重大挑战。针对以上问题,本文在深度学习框架下开展了面向多模态气象数据的短临降水学习方法研究。本文的主要研究内容和贡献归纳如下:

提出一种基于上下文信息的短临降水预测方法。该方法关注如何学习气象数据局部相关性。其核心思想是利用多模态气象数据的上下文信息,理解多模态气象数据在降水形成过程中的交互作用,并编码气象数据在空间和时间维度上的耦合关系。针对多模态特征的复杂耦合关系,该方法设计了一种模态加权融合策略,以在不同位置学习各个模态间的融合模式。在空间维度上,提出了一种多尺度时空特征编码策略,以增强模型对不同范围气象数据耦合关系的感知能力。此外,考虑到降水与多种气象要素间的相互影响,引入了降水数据调制多模态时空特征的建模过程,以编码时间维度上的耦合关系。在两个气象数据集上的实验结果证明了所提方法的有效性。

提出一种基于时空感知的短临降水预测方法。该方法关注如何建模气象数据时空特异性。其核心思想是利用时空信息,学习不同时空位置的气象特征编码方式。具体地,针对气象数据的时空特异性,该方法提出了一种时空感知卷积。在数据驱动的方式下,卷积核参数并非全局共享,而是基于时空位置信息构造的,从而编码具有时空特异性的气象特征。考虑到学习多模态特征相互作用的必要性,该方法提出了一种隐式的分层多模态特征交互策略,以建模多层级的气象特征间的耦合关系。在两个气象数据集上的实验证明了该方法的有效性。

提出一种基于分布感知的短临降水预测方法。该方法关注如何缓解长尾分布带来的负面影响。其核心思想是通过建模降水分布,减少头部样本对全局信息的影响,从而提升尾部特征鲁棒性。具体地,该方法在自注意力机制的框架中引入了一种降水分布感知策略。通过建模降水分布嵌入来指导区域嵌入的编码过程,增强了特征对不同降水区间的信息感知能力。建模分布嵌入与区域嵌入之间的相关性有助于在一定程度上缓解模型的偏向,进而增强尾部特征的判别性。此外,该方法还提出了一种邻域信息聚合机制,学习降水分布邻域内的非线性表示,从而进一步增强特征的表达能力。实验结果证明了该方法在短临降水预测任务上的有效性。

Other Abstract

Precipitation significantly affects various aspects of human life and production. Accurate precipitation nowcasting can offer crucial support for public services, including agricultural management, transportation planning, and disaster prevention. Despite its importance, it remains a challenging academic task. Recent advancements in deep learning have led to significant progress in weather prediction. However, most existing precipitation nowcasting methods based on deep learning employ radar echo data for prediction, overlooking the role of various meteorological elements in the  process. This dissertation studies precipitation nowcasting based on deep learning from multi-modal three-dimensional (height, longitude, and latitude) meteorological data. This research not only provides a novel modeling method and tools for precipitation nowcasting but also explores the feature encoding and nonlinear prediction of multi-modal spatiotemporal data under complex time-varying conditions. This has significant theoretical value and application prospects.

Given the complexity of the meteorological system, constructing a deep network for precipitation nowcasting based on multi-modal three-dimensional meteorological data will face several challenges. Firstly, meteorological data is a complex type of multi-modal spatiotemporal data, which generally has complex coupling relationships in multi-modal dimensions, spatial dimensions, and temporal dimensions. Therefore, how to capture the local relevance of meteorological data is the primary task of this paper.

Secondly, meteorological data would show regular variation in different regions and seasons, and this difference is particularly significant when the spatiotemporal positions are far apart. Therefore, how to model the spatiotemporal specificity of meteorological data is also a difficulty that needs to be solved.

In addition, precipitation data follows an extremely unbalanced long-tail distribution, which increases the difficulty of prediction of extreme precipitation. How to improve the prediction ability of tail classes under the long-tail distribution is a major challenge. Regarding these issues, this dissertation has carried out research on precipitation nowcasting methods for multi-modal meteorological data under the framework of deep learning. 

The main research contents and innovations are summarized as follows: A precipitation nowcasting method based on the context information is proposed. This method focuses on how to learn the local relevance of meteorological data. Its core idea is to employ context information of multi-modal meteorological data to capture the interaction of multi-modal meteorological data in the precipitation formation process and to encode the coupling relationship of meteorological data in spatial and temporal dimensions. In response to the complex coupling relationship of multi-modal features, this method designs a modality-weighted fusion strategy to learn the fusion pattern between each modality at different positions. In the spatial dimension, a multi-scale spatiotemporal feature encoding strategy is proposed to enhance the model’s perception ability of the coupling relationship of meteorological data in different ranges. In addition, to encode the coupling relationship in the temporal dimension, the precipitation data is introduced to guide the process of multi-modal spatiotemporal feature encoding, considering the mutual influence between precipitation and various meteorological elements. The experimental results on two meteorological datasets prove the effectiveness of the proposed method.

 

A precipitation nowcasting method based on spatiotemporal awareness is proposed. This method focuses on how to model the spatiotemporal specificity of meteorological data. Its core idea is to utilize spatiotemporal information to learn the encoding method of meteorological features at different spatiotemporal positions. Specifically, in response to the spatiotemporal specificity of meteorological data, this method proposes a spatiotemporal-aware convolution. In a data-driven manner, the convolution kernel parameters are not globally shared but are constructed based on spatiotemporal information, thereby encoding meteorological features with spatiotemporal specificity. Considering the necessity of learning the interaction of multi-modal features, this method proposes an implicit hierarchical multi-modal feature interaction strategy to model the coupling relationship between multi-level meteorological features. The experiments on two meteorological datasets have proven the effectiveness of this method.

A precipitation nowcasting method based on grouping learning is proposed. This method focuses on how to mitigate the negative impact of long-tail distribution. Its core idea is to model the distribution of precipitation to reduce the influence of head samples on global information, and thereby enhance the robustness of tail features. Specifically, this method introduces a precipitation distribution-aware strategy in the framework of self-attention mechanism. Modeling the embedding of precipitation distribution to guide the encoding process of patch embedding, enhances the feature’s information perception ability for different precipitation intervals.  Modeling the correlation between distribution embedding and patch embedding helps to alleviate the bias of the model to head classes, thereby enhancing the discriminability of tail features. In addition, this method also proposes a neighborhood information aggregation mechanism to learn the nonlinear representation within the neighborhood of the precipitation distribution, thereby further enhancing the feature’s expression ability. Experimental results prove the effectiveness of this method in precipitation nowcasting tasks.

Keyword深度学习 短临降水预测 多模态数据挖掘 时空特征建模 长尾学习
Subject Area模式识别
MOST Discipline Catalogue工学::控制科学与工程
Language中文
IS Representative Paper
Sub direction classification人工智能+科学
planning direction of the national heavy laboratoryAI For Science
Paper associated data
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/57515
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
靳淇兆. 基于深度学习的短临降水预测方法研究[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
基于深度学习的短临降水预测方法研究.pd(19432KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[靳淇兆]'s Articles
Baidu academic
Similar articles in Baidu academic
[靳淇兆]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[靳淇兆]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.