文献类型：期刊文献

中文题名：基于GA-Elman神经网络模型的年径流预测

英文题名：GA-Elman neural network model-based annual runoff prediction

作者：李志新 赖志琴 龙云墨

第一作者：李志新

机构：[1]贵州理工学院土木工程学院

第一机构：贵州理工学院土木工程学院

年份：2018

卷号：49

期号：8

起止页码：71-77

中文期刊名：水利水电技术

外文期刊名：Water Resources and Hydropower Engineering

收录：CSTPCD;;北大核心:【北大核心2017】；CSCD:【CSCD_E2017_2018】；

基金：贵州省科学技术基金计划(黔科合基础[2016]1062);国家自然科学基金项目(51508121);贵州省科技合作计划(黔科合LH字2016[7096])

语种：中文

中文关键词：遗传算法;神经网络;预测;模型

外文关键词：genetic algorithm;neural network;prediction;model

摘要：针对传统神经网络模型静态性及训练算法易陷入局部极值的缺陷,为了实现神经网络训练全局寻优,提高模拟精度,并使网络结构能动态反映年径流系列的时变特性,本文以年降雨及气温作为输入因子、年径流量为模型预测对象,结合遗传算法和Elman神经网络各自的优点,采用遗传算法对网络权值阈值全局优化,通过二者的耦合构建了GA-Elman年径流预测模型。利用构皮滩站1961—2015年的径流系列对模型进行了训练及测试,并对各模型预测性能比较分析。结果表明:GA-Elman模型预测平均相对误差5.29%、均方根误差55.81 mm,效果良好,对于径流预测具有实用价值;神经网络模型预测精度优于基于线性方法的模型,预测平均相对误差从12.01%降至7.07%以下;采用遗传算法改进神经网络权值阈值优化过程,预测平均相对误差从7.07%降低到5.29%,可明显提高模型泛化能力,从而改善径流预测效果。
Aiming at the static nature of the conventional neural network model and its defect of that the training algorithm is easy to fall into local extremum,the weights and thresholds of the network are globally optimized with genetic algorithm in combination with the merits of both genetic algorithm and Elman neural network and then the GA-Elman annual runoff prediction model is established herein through coupling both of them by taking the annual rainfall and temperature as the input factors and the annual runoff as the predicting object,so as to realize the global optimization of neural the network training for enhancing the relevant simulation accuracy,thus make the network structure has the dynamic performance that can reflect the time-variant characteristics of the annual runoff series. The model is trained and tested with the runoff series（ 1961 ~ 2015） at Goupitan Hydrological Station and then comparatively analyzed with the prediction performances of all the relevant models. The result shows that the prediction effect of GA-Elman model is better with the mean relative error of 5. 29% and the root mean square error of 55. 81 mm,thus has a practical value for runoff prediction. Moreover,the prediction accuracy of the neural network model is better than those of the linearized method-based models,of which the mean relative prediction error is decreased from 12. 01% to less than 7. 07%,while improving the optimization process of the neural network weight and threshold with genetic algorithm,the mean relative prediction error can be decreased from 7. 07% to 5. 29% and the generalization ability of the model can be significantly enhanced,thus the effect of runoff prediction can be improved as well.