SU Hui-dong1,2 ,JIA Yang-wen2 ,NI Guang-heng1 ,GONG Jia-guo2 , CAO Xue-jian1 ,ZHANG Ming-xi 1 ,NIU Cun-wen2 ,ZHANG Di 2
China Rural Water and Hydropower. 2018, (6):
40-43.
Download PDF
(
)
Knowledge map
 
Save
Machine learning has a wide range of applications in various fields and interdisciplinary subjects. In this paper,two machine learning methods,support vector regression SVR and BP neural networks,are used to learn,train and predict daily runoff based on the daily precipitation and runoff data of 2009 - 2014 in Yangtze River Basin. To compare with the traditional distributed hydrological model( THREW) ,The Ns efficiency coefficient and percent bais PB are adopted as indicators. The results show that the THREW model,with clear physical process and hydrological mechanism,has good simulation effect and the Ns coefficient efficiency is 0.503. However,the PB is big,
and the data and the operation is complex. The two machine learning methods work better in relative error index,and all have good runoff prediction and generalization ability. That is the ability to apply learning outcomes to new knowledge. But they are more dependent on data,
and the larger the data sample size is,the prediction results will be fitter. The prediction of runoff results by BP neural networks performs well in index,but it is distorted in the prediction of maximum runoff. The simulation results showe that the BP neural network is superior to the
THREW model while the THREW model is better than the SVR simulation results.