茅洲河流域降雨特征分析及其对径流与水质的影响研究

ANALYSIS OF RAINFALL CHARACTERISTICS AND THEIR INFLUENCE ON RUNOFF AND WATER QUALITY IN MAOZHOU RIVER

李易凡

11849073

硕士

环境科学与工程

刘俊国

2020-05-29

2020-07-10

哈尔滨工业大学

深圳

在气候变化与快速城市化背景下，中国城市河流健康面临巨大威胁。城市水循环过程复杂且受人类活动影响较大，同时长时间序列实测水文资料相对缺乏，导致城市河道径流模拟预测与水质分析较为困难。目前，基于物理机制的水文模型和神经网络模型是城市降雨径流模拟预测的重要工具，但分析总结这两类模型模拟特性的研究仍相对匮乏。此外了解降雨对径流污染的影响是城市河道水质管理的重要环节。茅洲河是深圳市第一大河，但近年来由于受到工业化与城市化的影响，导致该流域洪涝事件频发、河道水质污染问题严重。目前针对该流域降雨径流模型适用性及径流污染分析的研究仍然不足。针对这些问题，本研究以茅洲河流域为研究区域，首先分析总结了茅洲河流域的降雨时空分布规律及典型降雨特征。然后综合考虑资料获取情况与流域特点等因素，利用物理机制驱动的水文模型（SOBEK）和人工神经网络（ANN）、自适应模糊神经网络（ANFIS）与长短期记忆网络（LSTM）在茅洲河流域进行降雨径流建模，分析不同模型在水环境复杂的城市流域的性能差异，以及降雨特征对模型性能的影响。最后，利用数理统计和机器学习方法分析了降雨特征对径流污染的影响，并在此基础上建立降雨特征径流污染预测模型。本研究的主要内容及结论如下：一、降雨时空分布规律与典型降雨特征分析：茅洲河流域中小雨雨型比例较高，暴雨对总降雨量的贡献率最大，且降雨量近年呈一定的上升趋势。该流域降雨受地形等影响，空间分布不均，呈现出由东南向西北递减的趋势。通过对该流域3707个场次降雨事件进行统计分析，总结得出了不同雨型降雨特征（如降雨量、干旱期、降雨强度等）的分布与典型值。茅洲河流域前峰型降雨较为典型，随着雨量的增大均匀型与后峰型降雨逐渐增多；二、SOBEK模型与神经网络模型的对比：SOBEK模型在茅洲河流域展现了较强的适应性，整体NSE为0.891，场次降雨径流模拟的NSE值集中于0.7-0.8，但建模需大量基础数据，率定也需要较长时间。神经网络模型对于输入数据和模型的调试等要求远小于基于物理机制的水文模型，模拟效果与传统的水文模型具备可比性，甚至展现了一定的优势。ANN模型建模较为简易，整体模拟精度NSE为0.755，但由于模型本身特性限制，导致ANN模型预测稳定性较差。ANFIS 模型精度相对较好，NSE为0.864，但对于输入数据的处理有较高要求，处理大量数据时较为繁琐。LSTM模型模拟在响应的稳定性和洪峰预测性能上均表现出巨大优势，整体NSE可达0.945，展现其在城市降雨径流分析中的巨大潜力。此外本研究利用主成分分析方法，探究了场次降雨径流事件的降雨特征与模型性能的关系，发现这些模型存在一定的优势互补空间；三、降雨特征对水质污染的影响分析：本研究以降雨事件的COD平均浓度（EMC）作为水质污染的表征因子，在茅洲河一级支流石岩河进行分析，结果表明前峰型降雨和小雨的污染风险较大，干旱期和降雨强度是影响EMC较重要的降雨特征。通过构建的降雨特征EMC预测模型，发现在典型降雨特征情景下，小雨径流污染的EMC值约为中雨的2倍及大雨的5倍。此外还根据典型降雨特征下的EMC预测结果，对估算流域降雨径流污染年负荷方法进行了一定改进。本研究期为城市流域降雨径流分析和水质污染分析提供了一种研究思路，同时为城市雨洪管理、雨水资源化利用与河道水质管理提供一定参考依据。

In the context of climate change and rapid urbanization, the health of urban rivers in China is under great threat. The urban water cycle process is complicated and greatly affected by human activities. Meanwhile, the lack of measured hydrological data in a long time series leads to the difficulty in runoff prediction and water quality analysis of urban watercourses. At present, hydrological models based on physical mechanism and neural network models are important tools for urban rainfall runoff simulation and prediction. However, research on the analysis and summary of the simulation characteristics of these models are still relatively lacking.In addition, understanding the impact of rainfall on runoff pollution is an important part of urban river water quality management.The Maozhou River is the longest river in Shenzhen, but due to the impact of industrialization and urbanization in recent years, frequent flooding and waterlogging incidents have occurred in this basin, and the water quality of rivers has been seriously polluted.At present, researches on the applicability of rainfall runoff model and runoff pollution analysis are still insufficient.In response to these problems, this study takes the Maozhou River Basin as the research area. Firstly, it analyzes and summarizes the spatial and temporal distribution of rainfall and the typical rainfall characteristics of the Maozhou River Basin. Then comprehensively considering factors such as data acquisition and watershed characteristics, and using the physical mechanism-driven hydrological model (SOBEK) and Artificial Neural Network (ANN), Adaptive Network-based Fuzzy Inference System (ANFIS) network with Long Short-Term Memory (LSTM) in the Maozhou River Basin, analysis the performance of different models in complex urban river basin water environment, and the influence of rainfall characteristics on model performance. Finally, the influence of rainfall characteristics on runoff pollution is analyzed by means of mathematical statistics and machine learning, and the prediction model of runoff pollution is established on this basis.The main contents and conclusions of this study are as follows: 1. Analysis of the temporal and spatial distribution of rainfall and analysis of typical rainfall characteristics: The proportion of light rain and rain in the Maozhou River Basin is relatively high, the heavy rain contributes the most to the total rainfall, and the rainfall shows a certain upward trend in recent years. The uneven spatial distribution of rainfall in this basin is affected by terrain and other factors, showing a decreasing trend from southeast to northwest. Through statistical analysis of the 3707 rainfall events in the watershed, the distribution and typical values of different rain-type rainfall characteristics (such as rainfall, drought period, rainfall intensity and etc.) in the watershed are summarized. The front peak rainfall in the Maozhou River Basin is more typical, with the increase of rainfall depth, the uniform and post peak rainfall gradually increase; Second, the comparison between the SOBEK model and the neural network model: The SOBEK model shows a strong adaptability in the Maozhou River Basin, The overall NSE is 0.891, and the NSE values of the rainfall runoff simulations are concentrated at 0.7-0.8, but SOBEK modeling requires a lot of basic data, and it also takes a long time to determine. Neural network models require far less input data and model debugging than hydrological models based on physical mechanisms. The simulation results are comparable to traditional hydrological models, and even show certain advantages. ANN modeling is relatively simple, the overall simulation accuracy NSE is 0.755, but due to the limitations of the model itself, the prediction stability of the ANN model is poor. The accuracy of the ANFIS model is better. The NSE is 0.864, but it has higher requirements for the processing of input data, and it is more cumbersome when processing large amounts of data. The simulated response stability and flood peak prediction performance of LSTM have shown great advantages. The overall NSE can reach 0.945, demonstrating the great potential of the LSTM model in urban rainfall and runoff analysis. In addition, this study used the principal component analysis method to explore the relationship between the rainfall characteristics of the rainfall runoff events and the performance of the model, and found that these models have certain advantages and complementary spaces; Thirdly, the analysis of the impact of rainfall characteristics on water pollution: This study uses average concentration COD of rainfall events (EMC) as a characterization factor of water pollution, analysis of the Shiyan River, a first-level tributary of the Maozhou River, found that the above peak rainfall and light rain have a greater risk of pollution. The drought period and rainfall intensity are the most important rainfall characteristics that affect EMC . Through the construction of the rainfall characteristic EMC prediction model, it is found that under typical rainfall characteristic scenarios, the EMC value of small rain runoff pollution is about 2 times that of medium rain and 5 times that of heavy rain. In addition, according to the EMC prediction results under typical rainfall characteristics, the method of estimating the annual load of rainfall runoff pollution in the watershed has been improved.Through this study, it provides a research idea for the analysis of urban rainfall runoff and water quality pollution, and at the same time provides a certain reference for urban rainwater management, rainwater utilization and river water quality management.

降雨特征 水文模型 机器学习 径流模拟 水质分析

rainfall characteristics hydrological model machine learning runoff simulation water quality analysis

中文

联合培养

南方科技大学

李易凡. 茅洲河流域降雨特征分析及其对径流与水质的影响研究[D]. 深圳. 哈尔滨工业大学,2020.