基于支持向量机的石羊河流域径流模拟适用性评价

基于VC维和结构化风险最小理论的支持向量机方法因具有较好的学习和泛化能力而在预测预报领域得到广泛的应用。文中选取当月平均降水量、上月平均降水量以及当月平均相对湿度、平均最高气温和平均最低气温五个预报因子,采用Gridsearch算法优化参数,建立了基于支持向量机的月径流预报模型,并将其应用于石羊河流域八个子流域,定量分析其适用性。结果表明:模型在率定期和验证期模拟的平均Nash-Sutcliffe效率系数分别为0.831和0.806,相对误差分别在6%和5%以内;除个别峰值模拟较小之外,流量序列整体模拟效果较好;模型在丰水时段模拟值小于实测值,枯水时段模拟值大于实测值,在平水时段和枯水时段的模拟效果要优于丰水月份。因此,支持向量机模型在石羊河流域具有较好的适用性,可用于该流域的中长期水文预报。

Assessment on the applicability of support vector machine-based models for runoff simulation in Shiyang river basin

A monthly runoff forecasting models was proposed based on support vector machine(SVM) and was applied in eight sub basins of the Shiyang River basin.Because monthly runoff is associated with monthly precipitation at present month,the monthly precipitation of previous month,relative humidity,maximum temperature and minimum temperature at present month,so the five variables were selected as predictive factors,and Grid search algorithm was used to find optimal model parameters.After model was developed,applicability of the model was validated for the Shiyang River basin.The results indicate that the SVM model is satisfactory to represent complex non-linear relationships.During calibration and validation periods,the Nash-Sutcliffe coefficients were 0.831 and 0.806,and the relative error was controled under 6% and 5%,respectively.Although the peak flow was under-estimated to some degree,the simulated and observed runoff matches.Then,the runoff series was divided into three flow ranges: wet season(maximum 25%),normal season(middle 50%) and dry season(minimum 25%).Stream flow was under-estimated in wet seasons while over-estimated in dry seasons.Overall simulation in normal and dry seasons was better than that in wet seasons.Although the results obtained in this study may be a specific case,it can also reveal some possible issues in SVM model appliation and provide some helpful insights in the development and application of SVM models as well as other statistical learning methods in water resources management.