基于并行多尺度卷积记忆残差网络的物联网流量预测  

Research on Internet of Things Traffic Prediction Based on Parallel Multiscale Convolutional Memory Residual Network

作　　者：陆勤政 朱晓娟[1] Lu Qinzheng;Zhu Xiaojuan(Anhui University of Technology,Huainan 232001,China)

机构地区：[1]安徽理工大学,安徽淮南232001

出　　处：《廊坊师范学院学报（自然科学版）》2024年第1期33-41,共9页Journal of Langfang Normal University(Natural Science Edition)

基　　金：安徽省教育厅自然科学研究重点项目(KJ2020A0300)。

摘　　要：针对现有物联网流量预测方法中特征提取不足、丢失重要信息、预测准确度不高的问题,提出了一种基于并行多尺度卷积记忆残差网络的物联网流量预测方法。首先,采用并行结构,CNN提取多尺度的局部特征得到包含有局部特征的序列,LSTM和BiLSTM分别提取前向的时间关系和前后向的时间关系得到有合适比例的前后向时间特征序列;其次,引入ResNet结构,在CNN、LSTM、BiLSTM的输入和输出之间加入跳跃连接,即通过跳跃连接在特征序列中加入原始序列信息;再次,在有原始信息的特征序列中分配可训练的权重参数,突出相应序列的重要性,进行拼接得到总的输出序列;最后,将总的输出序列输入到全连接网络中得到预测结果。实验结果表明,本方法在均方根误差(RMSE)、平均绝对误差(MAE)、拟合系数(R2)3项指标上要优于其他方法,能更准确地进行物联网流量的预测。An accurate prediction of Internet of Things(IoT)traffic is crucial for the intelligent management of IoT systems,so based on a parallel multi-scale convolutional memory residual network,a prediction method is proposed to solve such issues as insufficient feature extraction,loss of crucial information,and low prediction accuracy.Firstly,a parallel structure is employed,where Convolutional Neural Networks(CNNs)with different kernel sizes extract multi-scale local features to obtain sequences containing local features.Long Short-Term Memory(LSTM)and Bidirectional LSTM(BiLSTM)are then utilized to capture forward and bidirectional temporal relationships,yielding properly proportioned forward and backward time feature sequences.Secondly,a Residual Network(ResNet)structure is introduced,incorporating skip connections between the input and output of CNN,LSTM,and BiLSTM,thereby integrating original sequence information through skip connections.Subsequently,trainable weight parameters are assigned to the feature sequences with original information,emphasizing the importance of respective sequences,followed by concatenation to obtain the overall output sequence.Finally,the total output sequence is fed into a fully connected network to derive the prediction results.The experimental results indicate that the proposed method outperforms existing approaches in terms of Root Mean Square Error(RMSE),Mean Absolute Error(MAE),and Coefficient of Determination(R2).More accurate predictions for IoT traffic can be achieved with this approach.

关 键 词：物联网流量预测 卷积神经网络 长短时记忆网络 双向长短时记忆网络 跳跃连接 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]