Rainfall-runoff Simulation Based on Runoff Classification Using Dynamic Artificial Neural Networks
Received date: 2011-01-17
Request revised date: 2011-06-02
Online published: 2012-01-20
Copyright
A runoff sequence is classified into several sub-runoff sequences with cluster analysis and practical application, and local dynamic neural network for each sub-runoff sequence is performed separately. An Elman recurrent neural network model (ENN) is constructed and applied for the daily runoff forecast in the Linyi sub-catchment of the upper Yishu river basin in this paper. In order to further evaluate the performance of local Elman neural network (LENN), global Elman neural network (GENN) is applied as a comparison at the same time in the study region. Based on analysis indexes such as Nash-Sutcliffe coefficient, correlation coefficient, mean relative error and root mean relative square error, the results of daily runoff and flooding processes are attained. It suggested that Elman recurrent neural network model has a high accuracy of forecast on the rainfall-runoff dynamic process and the maximum peak flow and peak occurrence time, whether it is configuration samples or evaluation samples. However, local Elman neural network based on runoff classification is more suitable and efficient to daily runoff forecasting than global Elman neural network, especially during the period of arid season and semi-humid season. The simulation precision of different periods decreases in the order of arid season, semi-humid season and humid season. Elman recurrent neural network model is feasible to effectively simulate the hydrological dynamic characteristic of the daily runoff and to reflect the complex nonlinear runoff regular pattern in the basin as a promising and efficient method. In order to further evaluate superiority of this method, longer series of data, more regions and hydrology model are need to study and analysis because of the highly nonlinear, spatial and temporal heterogeneity and dynamic uncertainty of rainfall-runoff process.
SHAO Yue-hong , LIN Bing-zhang , LIU Yong-he . Rainfall-runoff Simulation Based on Runoff Classification Using Dynamic Artificial Neural Networks[J]. SCIENTIA GEOGRAPHICA SINICA, 2012 , 32(1) : 74 -80 . DOI: 10.13249/j.cnki.sgs.2012.01.74
Fig. 1 Digital elevation model, the river net and meteorological stations of the Linyi sub-catchment图1 临沂流域DEM、数字河网及站点 |
Table 1 The runoff classification based on cluster analysis表1 聚类分析得到的径流分组 |
组别 | 第1组 | 第2组 | 第3组 |
---|---|---|---|
时间 | 11、12、1、2月 | 3、4、5、10月 | 6、7、8、9月 |
季节 | 枯水期 | 半湿润期 | 丰水期 |
代号 | LENN1 | LENN2 | LENN3 |
Fig.2 The Elman neural network architecture of rainfall-runoff simulation图2 降雨径流模拟的Elman神经网络结构 |
Table 2 Comparison of the simulation results from LENN and GENN of the runoff classification表2 基于径流分组的LENN和GENN模拟性能比较 |
样 本 | 分组 | LENN | GENN | ||||||
---|---|---|---|---|---|---|---|---|---|
DC | R | Wabs | RMSE | DC | R | Wabs | RMSE | ||
训练样本 | 枯水期 | 0.962 | 0.981 | 0.104 | 0.208 | 0.447 | 0.670 | 0.498 | 0.885 |
半湿润期 | 0.941 | 0.970 | 0.310 | 0.597 | 0.860 | 0.930 | 0.517 | 0.931 | |
丰水期 | 0.865 | 0.930 | 0.340 | 0.670 | 0.841 | 0.917 | 0.360 | 0.730 | |
全年 | 0.886 | 0.941 | 0.309 | 0.874 | 0.857 | 0.926 | 0.396 | 0.979 | |
检验样本 | 枯水期 | 0.856 | 0.925 | 0.161 | 0.261 | 0.481 | 0.693 | 0.348 | 0.524 |
半湿润期 | 0.719 | 0.845 | 0.261 | 0.441 | 0.175 | 0.418 | 0.555 | 0.985 | |
丰水期 | 0.781 | 0.884 | 0.363 | 0.549 | 0.757 | 0.870 | 0.373 | 0.583 | |
全年 | 0.820 | 0.905 | 0.332 | 0.972 | 0.792 | 0.890 | 0.393 | 1.050 |
Fig.3 The simulated results for daily rainfall-runoff from configuration samples in the Linyi sub-catchment in 2001-2005图3 临沂流域训练样本日径流量模拟结果(2001~2005) |
Fig.4 The simulated results for daily rainfall-runoff from evaluation samples in the Linyi sub-catchment in 2006-2008图4 临沂流域检验样本日径流量模拟结果 (2006~2008) |
Table 3 Relative error analysis for flooding process simulation in the Linyi sub-catchme表3 临沂流域次洪过程误差分析 |
样 本 | 次洪 编号 | Qmax (m3/s) | LENN | GENN | ||||
---|---|---|---|---|---|---|---|---|
模拟Qmax(m3/s) | 误差 (%) | Tpeak误差 (d) | 模拟Qmax(m3/s) | 误差 (%) | Tpeak 误差(d) | |||
训练样本 | 2001-7-31 | 2030.0 | 1504.4 | -25.89 | +1 | 1369.8 | -32.52 | +1 |
2002-7-25 | 138.0 | 134.4 | -2.61 | 0 | 143.2 | 3.77 | 0 | |
2003-9-7 | 1150.0 | 1044.2 | -9.20 | 0 | 934.9 | -18.70 | +1 | |
2004-8-5 | 1200.0 | 800.1 | -33.33 | +1 | 773.9 | -35.51 | +1 | |
2005-9-21 | 3760.0 | 3496.6 | -7.01 | 0 | 3364.1 | -10.53 | 0 | |
文本检验 | 2006-8-29 | 827.0 | 755.7 | -8.62 | +1 | 1155.8 | 39.76 | +1 |
2007-8-19 | 1334.6 | 858.8 | -35.65 | 0 | 1021.3 | -23.48 | 0 | |
2008-7-24 | 1814.4 | 1776.6 | -2.08 | 0 | 1547.0 | -14.74 | 0 |
The authors have declared that no competing interests exist.
[1] |
|
[2] |
|
[3] |
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
|
[9] |
|
[10] |
|
[11] |
|
[12] |
|
[13] |
|
[14] |
|
[15] |
飞思科技产品研发中心.神经网络理论与MATLAB7实现[M]. 北京: 电子工业出版社, 2005.
|
[16] |
|
/
〈 | 〉 |