【作者】 李逸；

【导师】 袁晓辉；

【作者基本信息】 华中科技大学 ， 水利工程， 2021， 硕士

【摘要】 水位预报是水文预报的一个重要组成部分,是水量调度的重要环节之一。由于水位受气候、人类活动、下垫面等多种不确定因素的影响,水位序列常常具有较大的不确定性。因此,及时准确的水位预报对水库调度、水资源配置等工作具有非常重要的作用。因此本文考虑水位序列的变化特性,选用骆马湖流域皂河闸水文站、洋河滩闸水文站、嶂山闸水文站、苗圩水文站、窑湾水文站、袁场水文站、张宅水文站、新店水文站、晓店水文站9个测站的月水位资料,结合Copula函数、灰色关联分析法(GRA)2种经验模态分解方法,经验模态分解1种数据处理方法以及长短期记忆神经网络(LSTM)、双向长短期记忆神经网络(Bi LSTM)、门控循环单元(GRU)3种预报模型,建立了16种月水位预报组合模型,进行水位预报研究,根据4个评定指标进行评价与优选,并在此基础上进行区间预报,本文取得的成果如下:(1)编制了Copula、Gray、EMD-Copula、EMD-Gray预报因子优选方案。将Copula函数、灰色关联分析法(GRA)结合经验模态分解方法(EMD)形成Copula、Gray、EMD-Copula、EMD-Gray四种预报因子筛选方法。由4个模型评定指标可以看出,基于经验模态分解的预报因子优选方案优于没有基于经验模态分解的预报因子优选方案,表明经验模态分解可以提高模型的预报精度。(2)构建长短期记忆神经网络(LSTM)、双向长短期记忆神经网络(Bi LSTM)、门控循环单元(GRU)、L-B-G耦合模型4种模型。由4个模型评定指标可以看出,基于L-B-G耦合模型的预报结果优于长短期记忆神经网络(LSTM)、双向长短期记忆神经网络(Bi LSTM)、门控循环单元的预报结果,表明耦合模型可以提高模型的预报精度。其中,在训练期,EMD-GRA-L-B-G模型的均方根误差为0.26米,绝对误差小于0.1米、0.2米、0.3米的占比分别为70.24%、82.14%、91.43%,在检验期,EMD-GRA-L-B-G模型的均方根误差为0.26米,绝对误差小于0.1米、0.2米、0.3米的占比分别为70.56%、83.33%、90.56%,预报效果最佳。(3)构建了适时区间预报。首先,计算总体预报时间在90%置信水平下的预报区间,用相应的指标对预报区间进行量化。然后对预报时间进行划分,对每个月分别进行区间预报,从而计算每个月分别在90%置信水平下的预报区间,简称适时区间预报。本文的结论可为骆马湖流域月水位的研究和应用做一种参考。更多还原

【Abstract】 Water level forecasting is an important part in hydrological forecast,and it plays a key role in water dispatching.Due to the influence of climate,human activities,underlaying surface etc,water level time series has great uncertainty.Therefore,promptly and accurately water level forecasting is significance for Reservoir Scheduling and water resources allocation.Taking Luoma Lake as the research object,this paper studies the influence of pre monthly water level on monthly water level prediction,and the achievements are as follows:(1)Four different factor optimization schemes were established.Copula function,grey relation analysis(GRA)and empirical mode decomposition(EMD)are used to form four factor optimization schemes,namely Copula,GRA,EMD-Copula and EMD-GRA.It can be seen from the four model evaluation indexes,the empirical mode decomposition(EMD)is better than the empirical mode decomposition(EMD)scheme,which indicates that EMD can improve the prediction accuracy of the model.(2)Long-short-term memory networks(LSTM),bidirectional long-short-term memory networks(Bi LSTM),gated-recurrent-unit(GRU)and the coupling model L-B-G were established.It can be seen from the four model evaluation indexes,coupling model is better than the single model in training and test period.The coupling model is effective as a prediction model to improve the accuracy and stability of the simulation.To be specific,in the training period,the root-mean-square error(RMSE)of EMD-GRA-L-B-G model is 0.26,and the proportion of absolute errors which are smaller than 0.1 m,0.2 m and 0.3 m respectively accounts for 70.24%,82.14% and 91.43% respectively.By contrast,in the test period,the RMSE of EMD-GRA-L-B-G model is 0.26 and the proportion of absolute errors which are smaller than 0.1 m,0.2 m and 0.3 m respectively account for 70.56%,83.33% and90.56%,reflecting the optimal prediction effect.(3)Interval prediction was also made for water level.First of all,calculate the prediction interval where the overall prediction time is below 90% level of confidence and use the corresponding indexes to quantify the prediction interval.Then divide the prediction time,and make interval prediction respectively per month so as to calculate the prediction interval under 90% level of confidence respectively per month(appropriate interval prediction).It was found through comparative analysis that the prediction quality of appropriate interval prediction is much better than that of single interval prediction and the appropriate interval prediction also stands out in terms of appropriateness.更多还原