【作者】 王浩；

【导师】 闫宝伟；

【作者基本信息】 华中科技大学 ， 水利工程， 2023， 硕士

【摘要】 径流预测是水文学的重要研究内容,是水库调度和流域水资源优化配置的重要依据。新时代水利高质量发展的需求,对径流预测提出了更高的要求,然而由于受水文、气象、下垫面等多个因素的影响,径流呈现出高度的非线性,加之气候变化和人类活动的不断加剧,传统的径流预测方法已很难适应新形势的发展要求。因此,充分利用现代技术手段,研究更加科学可靠的径流预测方法,成为水文学的重要任务之一。本文重点考虑径流形成过程中的时程相关性和空间相关性,尝试采用图卷积神经网络(GCN)进行径流预测模型的研究,主要研究内容与成果包括:(1)考虑径流序列的时程相关性,建立了基于长短时记忆神经网络(LSTM)的径流预测模型。以汉江上游20个干支流代表站点2000～2016年的实测日径流序列作为模型输入,将石泉、安康、白河、丹江口4个干流站点作为研究对象,分析了LSTM在汉江上游径流预测中的适用性。(2)针对LSTM对径流的空间相关性考虑不足的问题,基于GCN与门控循环单元(GRU),建立了时空图卷积神经网络(T-GCN)径流预测模型。利用GCN学习站点的拓扑结构以捕获径流的空间相关性,利用GRU学习径流序列的动态变化以捕获其时间相关性,从而综合考虑了径流的时空特征。将汉江上游20个代表站点的空间位置关系转化为邻接矩阵,与包含历史径流数据的特征矩阵一并作为模型输入,对建立的T-GCN进行模型训练,并与LSTM进行了对比分析。(3)针对T-GCN难以反映拓扑结构动态变化的问题,基于图卷积递归单元(GCRU)和时空元图学习器(STG),建立了元图卷积递归网络(Mega CRN)径流预测模型。STG作为一种新的时空数据图结构学习机制,可以解决拓扑结构时间和空间异质性的问题,将其嵌入GCRU编码-解码器,从而可以实现水文站点拓扑结构的动态反映。将MegaCRN径流预测模型应用于汉江上游,并与LSTM、T-GCN进行了对比分析。更多还原

【Abstract】 Runoff prediction is an important part of hydrology,and is an important basis for reservoir scheduling and optimal allocation of water resources in the basin.However,due to the influence of many factors such as hydrology,meteorology and substratum,runoff is highly non-linear,and due to the increasing climate change and human activities,the traditional runoff prediction methods can hardly adapt to the new development requirements.Therefore,it has become one of the important tasks of hydrology to make full use of modern technology to study more scientific and reliable runoff prediction methods.In this paper,we focus on the temporal and spatial correlations in the runoff formation process,and try to adopt graph convolutional neural network(GCN)for the research of runoff prediction model,the main research contents and results include:(1)A runoff prediction model based on long and short term memory neural network(LSTM)is established considering the temporal correlation of runoff sequences.The measured daily runoff series from 2000 to 2016 at 20 representative stations of the upper reaches of the Han River were used as model inputs,and four main stream stations,namely Shiquan,Ankang,Baihe and Danjiangkou,were taken as research objects to analyse the applicability of LSTM in runoff prediction of the upper reaches of the Han River.(2)To address the problem that the LSTM does not sufficiently consider the spatial correlation of runoff,a spatio-temporal graphical convolutional neural network(T-GCN)runoff prediction model is developed based on GCN and gated recurrent units(GRU).The GCN is used to learn the topology of stations to capture the spatial correlation of runoff,and the GRU is used to learn the dynamics of runoff sequences to capture their temporal correlation,thus taking into account the spatial and temporal characteristics of runoff.The spatial location relationships of the 20 representative stations in the upper reaches of the Han River were transformed into an adjacency matrix and used as model input together with a feature matrix containing historical runoff data,and the established T-GCN was trained as a model and compared with the LSTM for analysis.(3)A meta-graph convolutional recurrent network(MegaCRN)runoff prediction model was established based on graph convolutional recurrent units(GCRU)and spatiotemporal meta-graph learners(STG)to address the problem that the T-GCN can hardly reflect the dynamic changes of topological structure.STG,as a new mechanism for learning the graph structure of spatio-temporal data,can solve the problem of temporal and spatial heterogeneity of topological structure by embedding it into the GCRU coder-decoder,thus enabling dynamic reflection of hydrological station topology.The Mega CRN runoff prediction model is applied to the upper Han River and compared with LSTM and T-GCN.更多还原