基于CNN-BiGRU-Att融合模型的人民币汇率预测研究  

Research on RMB Exchange Rate Prediction Based on CNN BiGRU Attention Fusion Model

作　　者：徐伟 陈秀明 储天启 XUWei;CHEN Xiuming*;CHU Tianqi(School of Computer and Information Engineering,Fuyang Normal University,Fuyang 236037,China)

机构地区：[1]阜阳师范大学计算机与信息工程学院,安徽阜阳236037

出　　处：《安庆师范大学学报（自然科学版）》2023年第2期35-41,共7页Journal of Anqing Normal University(Natural Science Edition)

基　　金：安徽省哲学社会科学规划青年项目(AHSKQ2021D47);中国高校产学研创新基金(2019ITA01037);安徽省高校自然科学研究重点项目(KJ2019A0541)。

摘　　要：为了更好地预测人民币汇率的走向,本文提出了一种引入注意力机制(Attention)的卷积神经网络(CNN)和双向门控循环单元(BiGRU)的融合模型,对普通模型的单一模式适当加以改良,使之能够更好地反映汇率变动的走势。该模型使用一维卷积神经网络Conv1D来挖掘数据的内部特征,BiGRU获取数据间的时序特征,通过引入注意力机制,能够快速让模型获取更多关键的特征信息。选用平均绝对误差(MAE)、均方误差(MSE)以及平均绝对百分比误差(MAPE)作为评价指标进行比较实验,相比于其他模型评估指标体系的值,该模型结果均更优,验证了CNN-BiGRU-Att模型在汇率预测上的可行性。In order to predict the trend of RMB exchange rate better,a fusion model of convolutional neural network(CNN)and bidirectional gated recurrent unit(BiGRU)with attention mechanism is proposed to appropriately improve the sin-gle mode of the ordinary model in this paper to improve the prediction of RMB exchange rate trends.The model blends CNN’s one-dimensional convolutional neural network Conv1D to mine the internal features in the data,and BiGRU to obtain tempo-ral features between the data.Additionally,by introducing the Attention mechanism,the model can quickly obtain more key feature information.MAE,MSE and MAPE are selected as evaluation indexes for comparative experiments.Compared with other evaluation indexes,this model got a better result.The feasibility of CNN-BiGRU-Att model in exchange rate prediction is verified.

关 键 词：卷积神经网络 双向门控循环单元 注意力机制 汇率预测 

分 类 号：TP-391[自动化与计算机技术]