长江源高寒区域河川径流预测方法及其对比分析

为了更好地预测长江源高寒区域水文循环尤其是径流过程的变化规律,根据沱沱河站(1961年-2009年)共49年的降水和径流资料,以降水量作为输入向量,径流量作为目标向量,分别采用偏最小二乘回归估计、改进的BP神经网络和RBF神经网络建立了径流预测模型,并利用Matlab工具软件编程求解。通过三种计算方法预测结果的分析与对比表明:偏最小二乘回归估计模型的径流预测结果基本合理,但该方法需要降水数据作为已知条件,同时要求降水和径流的相关性较高,对于长江源高寒区域来水复杂的地区不是很合适;改进的BP网络模型因受到神经网络学习和训练的随机性影响,需要相当大的运算量,而且预测精度也不高,如果合理选择RBF神经网络模型的周期和spread值,其径流预测结果的精度相对较高,所以是值得推荐的方法。

Comparative Analysis on the Predictive Methods of Annual Runoff in the Souce Region of the Yangtze River

The source region of the Yangtze River,whose main function is to protect the long-term sustainability of the Yangtze River and to supply high-quality water to the downstream,is one of the most important areas of the Three-River headwaters region in China.In recent 50 years,the alpine ecosystem and permafrost were degraded continuously under climate change,the annual runoff decreased continuously as well,but in the meanwhile the frequency of runoff greater than 550 m3/s increased obviously.In order to predict the regional water cycle,especially the variation of runoff,in the source region of the Yangtze River,three different predictive runoff models are developed using the observed precipitation and runoff data from 1996 to 2009 at the Tuotuohe river station.The three models adopt three different methods,including the least-square regression,improved BP neural network,and RBF neural network,and Matlab is used to analyze the results.The predictive results of the least-square regression method are reasonable but this method requires a known precipitation and a high correlation between precipitation and runoff,which is not applicable for the source region of the Yangtze River due to its complex sources of incoming water.The improved BP neural network method requires large computational burden and lacks a precise prediction of runoff.The RBF neural network can provide a high prediction precision if the cycle and spread value of the model are selected reasonably.