基于融合神经网络的短期交通流预测研究  

Research on Short-term Traffic Flow Prediction Based on Fusing Neural Network

作　　者：李杰[1] 张子辰 孟凡熙 朱玮[1] LI Jie;ZHANG Zi-chen;MENG Fan-xi;ZHU Wei(School of Electronics and Control Engineering,Chang′an University,Xi′an710064,China)

机构地区：[1]长安大学电子与控制工程学院,西安710064

出　　处：《兰州交通大学学报》2022年第3期60-67,91,共9页Journal of Lanzhou Jiaotong University

摘　　要：针对交通流预测精度受到原始信号噪声、模态混叠等因素的影响,提出了一种基于集成经验模态分解和双向长短期记忆网络的融合神经网络模型,并加入注意力层,实现短期交通流预测.集成经验模态分解将交通流信号分解成多个固有模态分量,双向长短期记忆网络从正向和反向同时读取序列,并将输出的信号输入注意力层.实验结果表明:所提出的模型成功抑制了模态混叠现象,并且可以更充分地学习交通流序列中的时序特征,且注意力机制能够捕捉整个交通流时间序列更加有影响力的时间点,并合理分配其训练权重,提高递归模型的特征提取能力;所提出的集成经验模态分解-长短期记忆网络-注意力机制模型的平均绝对百分比误差和拟合优度分别为1.2318%和0.9410,优于其他6种竞争模型,在短期交通流预测中具有较高的精度和稳定性.Aiming at the influence of the original signal noise,modal aliasing and other factors on the prediction accuracy of traffic flow,a fusion neural network model based on ensemble empirical mode decomposition(EEMD)and bidirectional long short-term memory(BiLSTM)network is proposed,and an attention layer is added to realize the short-term traffic flow prediction.The EEMD decomposes the traffic flow signal into multiple intrinsic modal components,and the BiLSTM network reads the sequence from the forward and reverse simultaneously,and feeds the output signal to the attention layer.The experimental results show that the model can suppress the phenomenon of modal aliasing successfully,which will learn the time series features more fully in the traffic flow sequence,and the attention mechanism(AM)can capture the more influential time points of the entire traffic flow time series and allocate its training weights reasonably so as to improve the extraction ability of the recursive model.The mean absolute percentage error and goodness of fit of the EEMD-BiLSTM-AM model proposed in this paper are 1.2318%and 0.9410 respectively,which are better than other 6 competing models.The model has high accuracy and stability in traffic flow prediction.

关 键 词：智能交通 短期交通流预测 集成经验模态分解 双向长短期记忆网络 注意力机制 

分 类 号：U491.1[交通运输工程—交通运输规划与管理]