水利水电科技进展,2014,34(2)摇 Tel:025 83786335摇 E鄄mail:jz@ hhu. edu. cn摇 http: / / kkb. hhu. edu. cn
第 34 卷第 2 期
Vol. 34 No. 2
水 利 水 电 科 技 进 展
Advances in Science and Technology of Water Resources
2014 年 3 月
Mar. 2014
作者简介:崔东文(1978—),男,云南玉溪人,高级工程师,主要从事水资源水环境研究。 E鄄mail:cdwgr@ 163. com
DOI:10. 3880 / j. issn. 1006 7647. 2014. 02. 013
多重组合神经网络模型在年径流预测中的应用
崔东文
(云南省文山州水务局,云南 文山摇 663000)
摘要:针对线性组合预测模型预测精度不高、单一预测模型权重较难确定和非线性组合预测模型组
合函数难以构造等问题,为最大限度地挖掘输入向量间的有用信息以及充分发挥神经网络模型的
高度非线性映射能力,提出一种基于 BP、Elman、RBF、GRNN 这 4 种神经网络算法原理的多重组合
年径流预测模型。 以 4 种单一预测模型的预测结果作为一次组合预测模型的输入向量,实测流量
作为输出向量,构建 4 输入 1 输出的一次组合预测模型;再以一次组合预测模型预测结果作为二次
组合预测模型的输入向量,实测流量作为输出向量,构建 4 输入 1 输出的二次组合预测模型;依次
类推,构建 12 种多重组合预测模型。 以新疆伊犁河雅马渡站年径流预测为例,将预测结果与 4 种
单一预测模型及 IEA鄄BP 模型的预测结果进行比较,结果表明:多重组合预测模型的预测精度和泛
化能力较单一预测模型均有较大提高,随着模型组合重数的增加,预测精度呈提高趋势,是提高预
测精度的有效方法。
关键词:径流预测;组合模型;BP 神经网络;Elman 神经网络;RBF 神经网络;GRNN 神经网络
中图分类号:TV121摇 摇 摇 文献标志码:A摇 摇 摇 文章编号:1006 7647(2014)02 0059 05
Application of multiple combined neural network model in annual runoff prediction/ / CUI Dongwen (Wenshan Water
Conservancy Bureau of Yunnan Province, Wenshan 663000, China)
Abstract: In light of the problems that the prediction accuracy of a linear combination model is not high, the weight of a
single prediction model is hard to determine, and the function of a nonlinear combined prediction model is hard to
construct, a multiple combined annual runoff prediction model was put forward based on the principle of neural networks
such as BP, Elman, RBF, GRNN, so as to maximize the useful information between the input vectors, and give full play to
the characteristics of neural network models, such as their nonlinear mapping ability. Using the results of four single
prediction models as the input vectors of one combined prediction model, and the observed flow data as the output vector,
one combined prediction model with four inputs and one output was constructed. Then, using the result of the one combined
prediction model as the input vector for two combined prediction models, and the observed flow data as the output vector,
two combined prediction models with four inputs and one output were constructed. Using the same method, multiple
combined prediction models with 12 kinds of construction schemes were constructed. Taking the Yili River Yamadu
hydrological station in Xinjiang as an example, its annual runoff prediction results were compared with those of four kinds of
single BP models and IEA鄄BP models. The results show that the prediction accuracy and generalization ability of multiple
combined prediction models are improved as compared with the single prediction model, and with the increase of the
combined number of model, the prediction accuracy tends to be improved. Multiple combined prediction models can
improve the prediction accuracy.
Key words:runoff prediction; combined model; back鄄propagation neural network; Elman neural network; radial basis
function neural network; generalized regression neural network
1摇 研究背景
河川径流受多种因素的影响和制约,表现出复
杂、随机、多维等特性,探寻能够提高预测精度的模
型对河川径流预测预报具有重要的现实意义和应用
价值。 人工神经网络中常见的 BP 神经网络、Elman
神经网络、RBF 神经网络和 GRNN 神经网络均广泛
运用于径流预测
[1鄄5]
。 实践表明,不同的预测方法往
往有着不同的预测结果,不同的预测方法挖掘不同
的有用信息,不同模型的预测结果通常具有互补性,
其预测精度也各不相同
[6]
,任何一种预测方法都有
其适用性和局限性,应依据实际情况选择适当的模
型与方法
[7]
。 组合预测是 Bates 等
[8]
在 1969 年提
出的,其目的是为了有效地利用各种模型的优点,将
·95·