基于支持向量回归机的长期径流预报 及不确定性分析

根据汉江流域皇庄站1981-2008年逐月径流量与1980-2007年逐月74项环流指数、北太平洋海温场、500hPa高度场的相关关系,利用逐步回归挑选预报因子,构建基于遗传算法的支持向量回归机模型(GA-SVR),并对2009-2013年逐月径流量进行预报;结果表明,径流预报精度较高,汛期平均相对误差在30%以内,非汛期、年总量平均相对误差在20%以内,均优于随机森林和多元线性回归模型。将GA-SVR模型的预报结果作为概率预报的基础,采用贝叶斯理论中的水文不确定性处理器(HUP)对预报的可靠度进行分析;结果表明,HUP不仅可以提供精度更高的定值预报,还能以置信区间的方式量化预报的可靠度,提供更为丰富的预报信息。

Long-term runoff forecasting based on SVR model and its uncertainty analysis

In accordance with the Huangzhuang Station''s monthly runoff from 1981 to 2008 and the correlativity from 1980 to 2007 among the 74 circulation indexes of each month, the monthly north pacific sea surface temperature field, and the 500hPa geopotential height, we used the stepwise regression method to select the forecast factors and built a GA-SVR Model (Genetic Algorithm Support Vector Regression Model) on the basis of GA ( Genetic Algorithm) , inorder to forecast the monthly runoff from 2009 to 2013. The results showed that the accuracy of the runoff forecast was relatively high: the average relative error in flood season was within 25% ; the yearly runoff amount was within 20% in non-flood season. It was superior to Random Forest and Multiple Regression Model. With the forecast results of the GA-SVR Model as the basis of the probability forecast, we used the Hydrologic Uncertainty Processor (HUP) of the Bayesian Theory to analyze the forecast reliability. The outcome indicated that HUP could not only give a constant-value forecast with relatively high accuracy , but also quantify the forecast reliability in the form of a confidence interval to provide more forecast information.