基于门控单元循环神经网络的台风路径预测  

TYPHOON TRACK PREDICTION BASED ON GATED RECURRENT UNIT NEURAL NETWORK

作　　者：徐高扬 郑海涛[1] 黄国庆 吴凤波[3] Xu Gaoyang;Zheng Haitao;Huang Guoqing;Wu Fengbo(School of Mathematics,Southwest Jiaotong University,Chengdu 611756,Sichuan,China;School of Civil Engineering,Chongqing University,Chongqing 400044,China;School of Civil Engineering,Southwest Jiaotong University,Chengdu 611756,Sichuan,China)

机构地区：[1]西南交通大学数学学院,四川成都611756 [2]重庆大学土木工程学院,重庆400044 [3]西南交通大学土木工程学院,四川成都611756

出　　处：《计算机应用与软件》2019年第5期119-125,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(51778546;51578471);四川省青年基金项目(2016JQ0005)

摘　　要：传统的神经网络结构不能很好地处理序列问题。通过对历史台风数据库中的台风分类,提出基于门控单元网络的台风路径预测模型。利用历史台风的经纬度信息,分别用普通循环神经网络、长短时记忆网络和门控单元网络预测台风未来6小时位置信息。实验表明,在测试集上门控单元网络具有最小的平均绝对误差,能够有效提高路径预测精度,与稀疏循环神经网络预测方法相比,有更小的平均绝对误差。The traditional neural networks cannot deal well with sequence problem.Through typhoon classification in historical typhoon database,we proposed a typhoon track prediction model based on gated recurrent unit network.Using the longitude and latitude information of historical typhoons,the position information of typhoons in the next 6 hours was predicted by general cyclic neural network,long-term and short-term memory network and gated recurrent unit network,respectively.The experiments show that the gated recurrent unit network has the smallest average absolute error in the test set,and it can effectively improve the accuracy of track prediction.Compared with the sparse cyclic neural network prediction method,the average absolute error is smaller.

关 键 词：动态规整 相似度 长短时记忆网络 门控单元网络 路径预测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]