CASIA OpenIR  > 精密感知与控制研究中心  > 人工智能与机器学习
定量降水机器学习模型与方法研究
吴雅婧
Subtype博士
Thesis Advisor张文生
2021-05
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Name工学博士
Degree Discipline模式识别与智能系统
Keyword机器学习 定量降水估计 定量降水预报 深度图神经网络 时空图卷 积网络
Abstract
随着社会经济的快速发展,准确、及时地获取定量降水信息在国计民生中发挥着愈加重要的作用,已成为日常生活、农业发展、环境治理、电力系统等各领域科学决策的重要依据。降水受诸多因素影响,各因素之间相互作用、关联错综复杂,传统气象方法受限于表达能力,无法准确刻画多种降水因素间的复杂潜在关联。如何有效提升定量降水估计和预报的精准度,从而满足我国经济和社会发展对定量降水服务水平日益增长的需求,成为目前气象服务中亟待解决的关键问题。伴随着气象观测手段的不断革新,各种气象观测数据呈现指数级增长,为定量降水提供了丰富的信息和全面的视角。机器学习方法在复杂非线性关系建模方面的优势突出,与气象大数据相结合,可望为大幅度提升定量降水估计和预报性能开辟一条新途径,逐渐成为国内外气象研究的一个热点。
探索各种气象观测数据之间及其与天气现象间的潜在规律和内在联系是人工智能服务气象应用的核心。本文以机器学习模型和方法为切入点,从定量降水估计、临近定量降水预报和降水集合概率预报三个典型任务入手,深入研究面向定量降水场景的机器学习模型和方法,旨在提升现有定量降水模型的效果,为公众提供更优质的气象服务。本文的主要研究内容和贡献可归纳如下:

1. 提出一种基于深度图卷积回归网络(Graph Convolutional Regression Networks, GCRNs)的定量降水估计模型。考虑到雨场的空间信息对预报精度的影响,该模型统一建模雷达反射率到真实降水值的非线性映射关系以及雨场的空间相关性:首先,将原始的雷达雨量计数据结构化为图数据,进而将定量降水估计问题形式化为图卷积网络学习问题,以建模雨场中的空间相关性;其次,利用深度图卷积模型,在建模空间相关性的基础上,学习雷达反射率到真实降水值的复杂非线性映射;最后,考虑局部强降水事件,提出了多卷积机制,以保持局部降水的特性不被周围信号干扰,从而有效提升复杂天气情形下的降水估计效果。在中国气象局杭州降水数据上的实验结果表明:所提方法在均方根误差、平均绝对偏差、中位绝对误差以及相关系数等评价指标上均优于国际先进的对比方法。所提模型已服务于中国气象局“雷达专项”项目。

2. 提出了一种基于时空图卷积的定量降水预报模型(Inductive spatiotemporal Graph Convolutional Networks, InstGCNs)。该模型通过数据驱动的方式学习历史雷达数据与未来降水量之间复杂的映射关系,同时利用降水场中的时空相关性实现更为精准的多步定量降水预报:首先,模型将定量降水预报形式化为时空图学习问题,提出了基于子图序列结构的归纳式学习机制,改善了时空图预测模型中由直推式学习带来的应用限制,实现对任意位置的降水预测;在此基础上,在图结构化的过程中提出椭圆相关性的构图方式,比现有常规的建图方式更能匹配实际观测中空间相关性;最后,根据差分平稳假设,提出节点差分模块用于时序建模,有效地捕捉雨场序列中非平稳动态模式。在2014年到2015年杭州和上海天气雷达覆盖范围内的多个降水过程上进行了详细实验验证,结果表明InstGCNs在定量降水预报任务中效果优于国际主流时空学习方法。所提模型已服务于中国气象局“雷达专项”项目。

3. 提出了一种多分布混合的集合概率预报模型(Mixture Probabilistic Model, MPM)。该研究工作借鉴机器学习中的混合高斯密度估计方法,以应对定量降水预报中不确定性难以刻画以及概率分布难以拟合的难题:首先,该模型实现了删失偏位的单分布的混合,并将其用在降水概率预报中,使得拟合的分布更加灵活地适应复杂的大气变化;其次,提出一种利用狄利克雷(Dirichlet)分布估计权重参数的新方法并将其用在融合分布中,通过迭代求解的方式优化权重,充分考虑单分布模型和混合分布模型的贡献,有效地校正原始集合成员预报的偏差。分别在奥地利因斯布鲁克(Innsbruck)和中国华东地区的降水集合预报上进行了实验研究,结果表明:MPM优于国际上先进的集合预报统计后处理方法。所提集合预报的统计后处理方法和预报输出的执行流程已工程化实现,正在业务应用推广。


Other Abstract

With the rapid development of social economy, the accurate and timely acquisition of quantitative precipitation information plays an increasingly important role in the national economy and people's livelihood, and has become an important basis for scientific decision-making in various fields such as daily life, agricultural development, environmental governance, and power systems. Precipitation is affected by many factors, and the interactions and correlations among various factors are complicated. Traditional meteorological methods are limited by their expressive ability and cannot accurately portray the complex potential correlations among various precipitation factors. How to effectively improve the accuracy of quantitative precipitation estimation and forecasting In order to meet the increasing demand of my country's economic and social development for quantitative precipitation service levels, it has become a key challenge to be solved in current meteorological services. With the continuous innovation of meteorological observation methods, various meteorological observation data show an exponential increase, providing a comprehensive perspective and rich information for quantitative precipitation. Machine learning methods have outstanding advantages in modeling complex non-linear relationships. Combined with meteorological big data, it is expected to open up a new way to greatly improve the performance of quantitative precipitation estimation and forecasting, and it has gradually become a hot spot in meteorological research.

Exploring the underlying laws and internal relations among various meteorological observation data and with weather phenomena is the core of artificial intelligence services for meteorological applications. This dissertation uses machine learning models and methods as the breakthrough point, starting with three typical tasks of quantitative precipitation estimation, quantitative precipitation nowcasting, and ensemble probability forecasting. We in-depth study of machine learning models and methods for quantitative precipitation, aiming to improve the performance of existing quantitative precipitation models and  provide the public with better meteorological services. The main research content and contributions of this dissertation can be summarized as follows:
1. A gauge-based graph convolutional regression networks (GCRNs) learning is proposed to formulate the quantitative precipitation estimation. In view of the shortages of the existing radar in estimating precipitation and the influence of the spatial information on the prediction accuracy,  GCRNs can simultaneously model two kinds of relationships: 1) the nonlinear mapping between radar reflectivity and rainfall rate and 2) the spatial correlation of rain field. This work innovatively structures the original radar rain gauge data into graph data and then formalizes the quantitative precipitation estimation problem into a graph convolutional network learning problem to model the spatial correlation in the rain field.    First, the deep graph convolutional network is used to learn the complex nonlinear mapping from radar reflectivity to real precipitation value as well as model spatial correlation. Furthermore, considering the local heavy rainfall, a multi-convolutional mechanism is proposed to prevent the characteristics of local rainfall from being over-averaged and effectively improve the precipitation estimation performance under complex weather conditions. Experimental results on the Hangzhou precipitation data of the China Meteorological Administration show that the proposed model is superior to other methods in evaluation metrics including root mean square error, average absolute deviation, median absolute error, and correlation coefficient. Additionally, the proposed model has been used in the "Radar Special Project" of China Meteorological Administration.

2. A novel deep learning model, Inductive spatiotemporal Graph Convolutional Networks (InstGCNs),  is proposed for multi-step quantitative precipitation forecasts. In order to cope with the problem that the temporal and spatial characteristics in precipitation forecasting are complicated and difficult-to-use, InstGCNs can both learn a nonlinear mapping from historical radar reflectivity to future rain intensity and extract complex spatiotemporal representations simultaneously by the data-driven approach. Specifically,  the quantitative precipitation forecasting problem is formalized as a spatio-temporal graph learning problem. The inductive learning mechanism is introduced into the sub-graph sequence, which improves the application restrictions brought by direct learning in the spatio-temporal graph model,  and realizes the precipitation prediction at any location. Then, for the construction of the graph, the proposed special elliptic structure could be more suitable to model spatial dependency for rain field in actual observation. Next, a new Node level Differential Block (Node-DB) is introduced to tackle non-stationary temporal dependency. From 2014 to 2015, multiple precipitation processes within the coverage of Hangzhou and Shanghai weather radars are verified by detailed experiments.  The experiments show InstGCNs outperforms state-of-the-art for precipitation forecasting. The proposed model has been used in the "Radar Special Project" of China Meteorological Administration.

3. A multi-distribution Mixture Probabilistic Model (MPM), is proposed to quantify the uncertainty in precipitation forecasting as well as solve the difficulty in fitting the probability distribution. MPM  is a  machine learning model and inspired by the mixed Gaussian density estimation. The mixture of  CSG0 distribution is realized for probabilistic precipitation forecasting to well match the complicated atmospheric changes. A novel method for estimating weight parameters is proposed using the Dirichlet distribution. This new weight parameters estimation method considers both correlation between and independence of ensemble members to take full advantage of raw ensemble forecasts, capturing the skills of mixture model and individuals simultaneously. The error of the original ensemble forecast is corrected by the weight parameters. The proposed MPM is tested using Innsbruck ensemble precipitation data and precipitation ensemble forecast data in East China. Compared with state-of-the-art post-processing approaches, MPM shows improved performance for all verification scores. The results in both cases indicate the potential and effectiveness of MPM for precipitation ensemble forecasting. The statistical post-processing method of ensemble forecast and the execution flow of forecast output proposed in this model have been implemented in engineering and are being applied in pilot projects.

Pages130
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/44827
Collection精密感知与控制研究中心_人工智能与机器学习
Recommended Citation
GB/T 7714
吴雅婧. 定量降水机器学习模型与方法研究[D]. 中国科学院自动化研究所. 中国科学院大学,2021.
Files in This Item: Download All
File Name/Size DocType Version Access License
吴雅婧博士论文.pdf(11694KB)学位论文 开放获取CC BY-NC-SAView Download
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[吴雅婧]'s Articles
Baidu academic
Similar articles in Baidu academic
[吴雅婧]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[吴雅婧]'s Articles
Terms of Use
No data!
Social Bookmark/Share
File name: 吴雅婧博士论文.pdf
Format: Adobe PDF
All comments (0)
No comment.
 

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