基于深度学习和时空关联建模的降水数据地形订正方法  

A Terrain-Based Bias-Correction Method for Precipitation Data Based on Deep Learning and Spatiotemporal Correlation Modeling

作　　者：吴雪峰 陈逸智 王昌栋[3] 黄栋[1] WU Xuefeng;CHEN Yizhi;WANG Changdong;HUANG Dong(College of Mathematics and Informatics,South China Agricultural University,Guangzhou 510642,China;Guangdong Meteorological Data Center,Guangzhou 51010,China;School of Computer Science and Engineering,Sun Yat-sen University,Guangzhou 510006,China)

机构地区：[1]华南农业大学数学与信息学院,广州510642 [2]广东省气象数据中心,广州510610 [3]中山大学计算机学院,广州510006

出　　处：《太原理工大学学报》2024年第4期734-742,共9页Journal of Taiyuan University of Technology

基　　金：国家自然科学基金资助项目(61976097);广东省气象局科学技术研究资助项目(GRMC2022Q05);广东省气象局协同观测和多源实况数据融合分析技术创新团队资助项目(GRMCTD202103)。

摘　　要：【目的】如何基于地形因素对降水数据进行误差订正(简称“地形订正”)是气象大数据研究的一个重要问题。但是,现有的降水数据地形订正方法存在两方面局限性:一是现有方法大多基于数据统计或传统机器学习而设计,未能拓展至特征学习能力更强的深度学习模型;二是现有方法亦未能结合气象数据的时间与空间关联信息以提升订正效果。【方法】提出一种基于深度学习与时空关联建模的降水数据地形订正方法。该方法首先对广东省降水数据和多源地形信息作预处理,结合经纬度进行对齐与编码,以构建相应的数据矩阵。进而结合各个格点的空间和时间邻域下的降水与地形数据进行多源时空信息建模,并结合降采样策略以缓解降水格点与非降水格点的数据不平衡性,最后构建深度神经网络进行回归订正。实验在广东省逾4000个气象观测站点、逾15万气象格点的真实气象数据上进行。【结果】实验结果验证了时空关联建模策略对提升降水订正效果的显著作用以及所提出方法相比于其他对比方法的性能优势。【Purposes】How to perform bias-correction for precipitation data based on terrain information has been an important problem in meteorological big data research.However,the existing precipitation terrain-based bias-correction methods mostly suffer from two limitations.First,most of them are designed based on statistical models or conventional machine learning models,thus fails to go beyond to the deep learning models with more powerful feature learning ability.Second,they also lack the ability to incorporate the spatiotemporal correlation information in meteorological data for enhancing the bias-correction quality.【Methods】To address these limitations,in this paper,a terrain-based bias-correction method for precipitation data based on deep learning and spatiotemporal correlation modeling is presented.First,the precipitation data and multi-source terrain information with alignment and encoding based on the longitude and latitude are preprocessed,and thus the corresponding data matrices are constructed.Then the precipitation and terrain data in the spatial and temporal neighborhood of each grid are adopted to achieve the multi-source spatiotemporal information modeling,and a down-sampling strategy is used to alleviate the imbalance between the precipitation grids and the non-precipitation grids.Finally,the deep neural network is constructed for regression and bias-correction.The experiments are conducted on the real meteorological data from over 4000 meteorological observation stations and over 150 thousand meteorological grids in Guangdong Province.【Findings】Experimental results have verified the significance influence of the spatiotemporal modeling strategy over precipitation bias correction quality and performance advantage of the proposed method over the baseline methods.

关 键 词：气象大数据 深度学习 降水量 地形订正 时空信息 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] S161.6[自动化与计算机技术—控制科学与工程]