神经网络算法在径流模拟中的开发与应用

Development and Application Of Neural Network Algorithm in Runoff Simulation

王萌

11749246

硕士

环境科学与工程

刘俊国

2019-05-29

2019-07-11

哈尔滨工业大学

深圳

在水文学研究中，过程驱动的水文模型受制于现实因素，可用输入数据和 模型所需数据往往不能完美契合，在一定程度上限制了其应用。而基于数据驱 动的神经网络模型具有良好的模型结构可变性和优异的性能，受到广大研究学 者的欢迎。在“降水-径流”关系模拟中，径流可能受到过去的气象因素的影响， 为了解决此问题，目前均采用了加长模拟时间步的办法。主流采用的人工神经 网络(ANN)模型拥有极强的学习能力，但不能体现出历史水文气象因素对未 来径流的影响力会逐步减弱这一事实。这不仅造成了模型泛化能力下降，而且 违背了自然规律。为了解决这一问题并测试其模拟性能以及最佳适用条件，本 研究应用了一种改进的神经网络——长短期记忆(LSTM)神经网络，在中国黑 河上游流域的不同时间尺度进行了“降水-径流”模拟。通过与 ANN 模拟结果 对比发现，LSTM 神经网络径流模拟的纳什系数(NSE)可达 0.8054，稍优于 ANN 的 NSE 值 0.7843;但是月尺度模拟方面，ANN 和 LSTM 的 NSE 值分别 可达 0.9077 和 0.8421。综合考虑模型泛化能力和鲁棒性，LSTM 更适合日尺度 模拟而 ANN 更适合月尺度模拟。此外，本研究还将神经网络模型与另一种基于 过程驱动的水文模型(FLEX-Topo)进行了比较，发现神经网络模型也同样具有 竞争力，纳什系数最大相差不过 0.048。基于前文 ANN 更适用于对月尺度模拟 的结论，本研究最后在月尺度上使用了 ANN 模型对未来多种气候变化的情景 下径流不同的变化趋势进行了预估。研究发现在温度和降水增加的情况下，未 来黑河上游水资源分布的不确定性会增加，这也为未来气候变化情景下的水资 源管理提供了参考依据。

In the current hydrological research, the process-driven hydrological models are limited by reality, the input data required and obtained cannot match perfectly, which limits the application to a certain degree. However, the data-driven neural network models have good flexibility and excellent performance, which is warmly welcomed by the researchers. In the “Rainfall-Runoff” simulation, the runoff may be affected by the meteorological factors in the past. The method of lengthening simulation time step is adopted to solve this problem in the current simulation. The mainstream Artificial Neural Network (ANN) model has strong learning ability, but it cannot reproduce the fact that the influence of historical hydrological and meteorological events on future runoff will gradually decline, making the generalization ability of the model decline and obviously violates the laws of nature. In order to solve this problem, test its performance and find the best suitable application conditions, an improved neural network structure, Long-Short-Term-Memory (LSTM) neural network is applied on the upper Heihe River Basin in China, to make the modeling for the precipitation-runoff relationship on different time scales. For the daily modeling, the Nash-Sutcliffe Efficiency Coefficient (NSE) of LSTM is 0.8054, which is higher than the NSE of ANN for 0.7843. But in the monthly modeling, the NSE of ANN and LSTM are 0.9077 and 0.8421 respectively. Taking the generalization ability and robustness into consideration, the LSTM is more suitable for daily modeling while ANN is a better option for the monthly simulation. In addition, this study also compares the two neural network models with another process-driven hydrologic model (FLEX-Topo) and it is found that the NSE gaps among them are marginal, which proves that the network models are competitive. Finally, based on the conclusion that ANN is more suitable for the monthly scale simulation, this study uses the ANN model to estimate the different change trends of runoff under various future climate change scenarios. It is found that with the increase of temperature and precipitation, the uncertainty of water resources distribution in the upper Heihe River Basin will increase in the future, which also provides references for water resources management in the future climate change scenario.

水资源管理 径流模拟 神经网络模型 水文模型 未来水资源变化

water resources management, runoff simulation, neuralnetwork models, hydrological models, future water resources change

中文

联合培养

南方科技大学

王萌. 神经网络算法在径流模拟中的开发与应用[D]. 深圳. 哈尔滨工业大学,2019.