基于多因子量化指标的支持向量机径流预测  

Prediction of runoff based on the multiple quantity index of SVM

作　　者：张楠[1,2] 夏自强[1,2] 江红[3] 

机构地区：[1]河海大学水文水资源与水利工程科学国家重点实验室,江苏南京210098 [2]河海大学水文水资源学院,江苏南京210098 [3]黄河流域水资源保护局,河南郑州450004

出　　处：《水利学报》2010年第11期1318-1324,共7页Journal of Hydraulic Engineering

基　　金：国家自然科学基金重大项目资助课题(30490235)

摘　　要：利用最小二乘支持向量方法,构建了基于多因子量化指标的径流预测模型。对长江上游寸滩水文站1981—2000年逐月蒸发量、水库容积指标量化后,以不同降雨量和气温作为输入量,建立了15种验证方案,通过穷举搜索二维最小化Gridsearch算法优化出惩罚因子和核宽度,并对2001—2006年逐月径流量进行预测。经验证,方案3、14和12精度均令人满意,其中,方案3精度最高(均方根相对误差为0.11,相关系数为0.89,确定性系数为0.88,输入影响因子为平均降雨量、平均气温、水库库容和蒸发量量化指标4项)。通过15种方案的比较发现,Qmax、Qmin、Tmin和Tmax对预测精度有弱化的作用,库容量化指标对预测精度的影响比蒸发量化指标Ezf大。基于多因子量化指标评价体系的支持向量机径流量预测模型体现了不同影响因子对径流量影响的相对程度,实现了精度和实用性的统一,为缺资料地区研究预报提供了新的方法。A runoff prediction model was developed based on the muhiple quantity index of SVM（support vector machine） method. According to the evaluation index of reservoir volume and evaporation, different temperature and rainfall parameters can be tested in the model training processes. The data at Cuntan Gauge Station in the period from 1981 to 2000 in the Upper Yangtze River were used for the model training, and 15 schemes were established. Grid search algorithm was used to find optimal regularization factors and kernel bandwidth, and to forecast the monthly runoff in 2001-2006 . The results indicate that scheme 3, 14, 12 are in high precisions, and that of scheme 3 is the highest（RMSRE 0.11, R^2 0.89, IA 0.88, inputs of this scheme are predicted evaluation index of average temperature, reservoir volume, average rainfall and evaporation）. The comparison of these 15 schemes shows that Q Qmin , Tmin , Tmax will reduce the precision of forecast, the influence of Vkr is larger than Ezf. The evaluation index SVM forecasting model which is based on influence factors of multi-plans, the unification of realiges precision and usability, provides a new forecast method to the lack-data watershed .

关 键 词：量化指标 径流量 支持向量机 Gridsearch 惩罚因子 核宽度 

分 类 号：P338[天文地球—水文科学]