【作者】 王伟；

【导师】 金保明；

【作者基本信息】 福州大学 ， 水利工程（专业学位）， 2020， 硕士

【摘要】 流域洪水预报是一种十分重要的防洪非工程措施,特别是对山区流域的洪水预报预警发挥着突出的作用。但是由于降雨时空分布不均匀,模型的系统误差,洪水分滞作用的影响等原因,导致预报结果出现误差。而这些误差的存在会影响预报结果在防洪减灾等方面的应用。因此,对于洪水预报误差分析研究就显得格外重要。论文首先根据崇阳溪上游流域的6个雨量站雨量资料和武夷山水文站前期流量资料,构建LMBP神经网络洪水预报模型,预报研究流域出口处的流量过程。接着,通过生成的预报误差,并根据极大熵原理推求误差的各阶概率密度函数,最终确定预报误差服从的具体分布规律。主要内容如下:（1）利用DEM（数字高程模型）提取研究区域的范围,并在ARCGIS平台上对流域进行划分,确定岭阳站、岚谷站、坑口站、大安站、吴边站、洋庄站6个雨量站控制流域面积的权重分别为0.16、0.17、0.15、0.18、0.22、0.12。（2）选取武夷山水文站21场洪水资料,其中13场作为训练样本,8场作为检验样本,建立崇阳溪上游流域LMBP洪水预报模型。模型的输入为经过归一化和赋权处理的6个雨量站雨量数据和武夷山站前期流量数据,输出为武夷山站预报流量数据,隐含层节点数经过试算采用8个,网络结构为3层。经过分析洪水流量过程预报相对误差的平均值均小于20%,洪峰流量相对误差均小于10%。（3）依据常见的极大熵模型约束条件的函数形式组成集合U,利用极差准则、方差准则、变异系数准则三个标准从集合U中选择出最合适的极大熵模型约束条件的函数形式为幂函数。（4）采用遗传算法对极大熵模型的参数λ_i进行求解,进而获得极大熵各阶概率密度函数的表达式。将洪水预报误差数据的概率密度直方图与极大熵各阶概率密度函数曲线进行拟合,根据图形拟合情况定性分析得到极大熵四阶概率密度函数优于极大熵二阶、三阶、五阶概率密度函数曲线。（5）运用Nash-Sutcliffe确定性系数E、理论——经验累计概率均方误差R_MSE、相关系数R²三个统计量,并绘制各阶理论——经验累计概率相关图,定量分析得到遗传算法求解的四阶概率密度函数为遗传算法求解的各阶函数中最优。（6）利用牛顿迭代法求解三阶、四阶概率密度函数,并与遗传算法求解的四阶概率密度函数通过三个统计量和各阶理论——经验累计概率相关图比较,最终确定牛顿迭代法求解的四阶概率密度函数为各条曲线中最优。将牛顿迭代法求解的四阶概率密度函数进行积分,进而确定预报流量误差在各个小区间发生的概率,为预报结果风险分析与防洪调度决策提供参考依据。更多还原

【Abstract】 Watershed flood forecasting is a very important non-engineering measure for flood control,especially for mountain torrents forecasting and early warning in mountain basins.However,due to uneven temporal and spatial distribution of rainfall,the systematic error of the model,the influence of flood lag and other reason,the forecast result have errors.The existence of these errors will affect the application of forecast results in flood control and disaster reduction.Therefore,it is very important to analyze and study the error of flood forecast.In this paper,based on the rainfall data of six rainfall stations in the upper reaches of Chongyang River and the discharge data of Wuyishan hydrological station in the previous hour,a LMBP neural network flood forecasting model is constructed to forecast the discharge data at the outlet of the basin.At the same time,through the generated prediction error,and according to the maximum entropy principle,the probability density function of each order of the error is deduced,and then the specific distribution law of the prediction error is determined.The main contents are as follows:(1)Using DEM(Digital elevation Model)to extract the scope of the research area,and divide the watershed on the ARCGIS platform.The area weights of Lingyang station,Langu station,Kengkou station,Daan station,Wubian station and Yangzhuang station are 0.16,0.17,0.15,0.18,0.22,0.12 respectively.(2)A total of 21flood data were collected,of which 13 were used as training samples and 8 as test samples.The input of the model is the normalized and weighted rainfall data of six rainfall stations and the discharge data of the previous hour of Wuyinshan station,and the output is the forecast discharge data of Wuyishan station,After trial calculation,8 points are used in the hidden layer structure,and the network structure is 3 layers.The forecast result is that the average relative error of flood flow process forecast error is less than 20%,and the average relative error of flood peak flow is less than 10%,which is in line with the standard accuracy.(3)The set U is formed by the functional form of the common constraint conditions of the maximum entropy model.The function form of the most suitable constraint condition of the maximum entropy model is the power function by using the range criterion,the variance criterion and the coefficient of variation criterion.(4)The genetic algorithm is used to solve the parameterλ_i of the maximum entropy model,and then the expression of the probability function of each order of the maximum tropy is obtanined.The probability density histogram of flood forecasting error data is fitted with the probability density function curve of each order of maximum entropy.According to the qualitative analysis of graph fitting,it is found that the fourth-order probability density function of maximum entropy is better than the second-order,third-order and fifth-order probability density function curve of maximum entropy.(5)Using the three statistics of Nash-Sutcliffe certainy cofficient E,theory-emprical cumulative probability mean square error and correlation coefficient.The cumulative probability correlation diagram of each order theory-experience is drawn,and the quantitative analysis shows that the fourth-order probability density function solved by genetic algorithm is the best among all order functions sovle by genetic algotithn.(6)The third-order and fourth-order probability density functions are solved by Newton iterative method,and compared with the fourth-order probability density functions solved by genetic algorithm through three statistics and empirical cumulative probability correlation diagrams.Finally,it is determined that the fourh-order probability density function solved by Newton iterative method is the best in each curve.The fourth-order probability density function solved by Newton iterative method is integrated to determine the probability of forecast flow error among each cell,which provides a reference basis for risk analysis of forecast results and decision-making of control operation.更多还原