河流健康评价的回归支持向量机模型及应用

建立河流健康评价指标体系、分级标准及回归支持向量机（ SVR ）河流健康评价模型，并以云南省文山州清水河健康评价为例进行研究。首先，利用层次分析法（ AHP ）从水文水资源、物理结构、水质、水生生物和社会服务功能5个方面遴选出13个评价指标，构建3个层次的河流健康评价指标体系和5个等级的分级标准；其次，基于SVR原理，利用随机生成和随机选取的方法，在等级标准阈值间构造5种不同容量大小的训练样本和检验样本，提出5种不同容量方案的SVR河流健康评价模型，设计合理的输出模式，并构建具有良好性能的RBF（radial basis function neural network ）回归模型作为对比模型，利用模型随机5次运行的平均相对误差绝对值、最大相对误差绝对值和运行时间对各方案模型性能进行评价；最后，利用达到期望精度的SVR模型对实例进行评价分析。结果表明：①无论是训练样本还是检验样本，5种方案的SVR模型的预测精度和泛化能力均优于 RBF模型。在相同参数设置条件下，SVR模型随着样本容量的增加其精度和泛化能力变化不大；而RBF模型随着样本容量的增加其精度和泛化能力均有提高。表明SVR模型具有较高的精度和泛化能力，可以用于河流健康评价，尤其在小样本情况下，SVR模型的精度和泛化能力是RBF模型不可比拟的。②5种方案的SVR模型对清水河2011-2012年3次调查的评价结果均为健康，但已接近于亚健康。

Regression support vector machine for river health assessment and its application

A river health assessment index system , grading standards , and a support vector regression （SVR） river health assessment model are proposed for health assessment of the Qingshui River in Wenshan , in Yunnan Province.In this study, first, 13 evaluation indices were selected with the analytic hierarchy process （AHP）in terms of hydrology and water resources , physical structure , water quality , aquatic organisms , and social services , in order to construct a three-level river health assessment index system as well as five -level grading standards . Then , based on the SVR principle , the random generation and random selection methods were used to construct five training and testing samples with different capacities in grading thresholds .Five models with different capacity solutions were developed for the SVR river ’ s health assessment .A reasonable output mode was designed , and the corresponding radial basis function neural network （ RBF） regression model , which showed a good performance , was built as a comparison model .After the model ran five times stochastically , the absolute value of the average relative error, the absolute value of the maximum relative error , and the runtime were used to evaluate the performance of the model in each program .Finally, the SVR model that achieved the desired accuracy was evaluated and analyzed in a case study .The results are as follows：（1） For either the training sample or the testing sample , the SVR model in five programs had a higher prediction accuracy and better generalization ability than the RBF model .Under the same parameter setting conditions , as the sample size increased , the SVR model ’ s accuracy and generalization ability changed insignificantly , while the RBF model ’s accuracy and generalization ability improved, indicating that the SVR model has higher accuracy and better generalization ability and can be used for river health assessment , especially in the cases of small samples.In this regard, the RBF model is totally unco