[目的]洪水分类预报能有效提高洪水预报准确性，为防灾减灾工作提供科学依据。[方法]针对分类因子和分类算法优选问题，以海河流域徒骇河宫家闸上游为例进行研究，(1)充分考虑产汇流影响因素与洪水特征，选取洪峰流量、洪水总量、时段洪量、洪水历时、起历时、落历时、峰度、偏度、涨水仰角、落水仰角、Cs、Cv、前3d面雨量、前10d面雨量、累计面雨量及最大面雨量等16维分类因子，使用主成分投影法(Principal Component Analysis，PCA)对分类因子降维提高计算效率;(2)基于密度聚类(Density-Based Spatial Clustering of Applications with Noise，DBSCAN)-随机森林(Random Forest，RF)算法进行洪水分类，减少对分类先验知识的依赖，提高了分类精度;(3)在徒骇河流域进行了方法应用，选择适用于半干旱半湿润地区的超渗-蓄满同时作用的产流模型及单位线汇流模型进行洪水分类预报研究，分别针对各类洪水进行模型率定。[结果]结果表明:轮廓系数为0。7015，表明DBSCAN算法聚类效果理想，基于RF算法的洪水分类准确率为91。67％，分类效果理想;经洪水分类预报，NSE系数均高于0。8，分类预报结果优于直接预报。[结论]结果说明:基于DBSCAN-RF洪水分类的洪水预报能较好地反映研究区域洪水演进过程，为研究区域洪水预报及防灾减灾工作提供依据。
Application research on flood classification-prediction based on DBSCAN-RF
[Objective]Flood classification-prediction can effectively improve the accuracy of flood prediction,thereby providing a scientific basis for flood disaster prevention and reduction.[Methods]In regards to the optimization of classification factors and classification algorithms,the upstream of Gongjia Gate of Tuhai River was took as an example.(1)Fully considering character-istics and influencing factors of historic flood,16 dimensional classification factors were selected,including peak flow,total flood volume,period flood volume,flood duration,duration of rising,duration of falling,kurtosis,skewness,elevation of rising,ele-vation of falling,Cs,Cv,surface rainfall in the first 3 days,surface rainfall in the first 10 days,cumulative surface rainfall,and maximum surface rainfall.Principal Component Analysis(PCA)was used to reduce the dimensionality of classification factors and improve computational efficiency.(2)Flood classification was carried out based on the Density Based Spatial Clustering of Applications with Noise(DBSCAN)-Random Forest(RF)algorithm to reduce the dependence on prior knowledge and improve classification accuracy.(3)The method was applied in the Tuhai River Basin.The runoff generation model that simultaneously reflects excess-infiltration and excess-storage,as well as the unit hydrograph confluence model,which are suitable for semi-arid and semi-humid areas,were chosen for flood prediction research.Model calibration was carried out separately for various types of floods.[Results]The result show that:(1)The contour coefficient was 0.701 5,indicating that DBSCAN algorithm had an ide-al clustering performance.The accuracy of flood classification was 91.67％,indicating that RF algorithm had an ideal classifica-tion performance;(2)The NSE coefficients of flood prediction were all higher than 0.8,and the result were better than direct prediction.[Conclusion]This indicates that flood classification-prediction based on DBSCAN-RF can better reflect the flood process in the research area,providing a basis for regional flood prediction and disaster prevention.
万方数据
维普
