计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (22): 284-290.DOI: 10.3778/j.issn.1002-8331.2105-0022

• 工程与应用 • 上一篇    下一篇

日交通流预测的编码器-解码器深度学习模型研究

曹阳,茅一波,施佺   

  1. 1.南通大学 信息科学技术学院,江苏 南通 226019
    2.南通大学 交通与土木工程学院,江苏 南通 226019
  • 出版日期:2022-11-15 发布日期:2022-11-15

Research on Encoder-Decoder Deep Learning Model for Daily Traffic Flow Prediction

CAO Yang, MAO Yibo, SHI Quan   

  1. 1.School of Information Science and Technology, Nantong University, Nantong, Jiangsu 226019, China
    2.School of Transportation and Civil Engineering, Nantong University, Nantong, Jiangsu 226019, China
  • Online:2022-11-15 Published:2022-11-15

摘要: 精准的日交通流预测是智能交通领域的重要研究内容之一。目前已有的日交通流预测模型大多在短期预测模型的基础上通过多步预测或者多目标预测的方式改进而来。这两种改进方案中,前者对误差的传播更为敏感,而后者则忽视了预测结果的时序关系,导致预测模型精度偏低。提出了一种用于日交通流预测的编码器-解码器深度学习模型,首先将长短时记忆网络(long short-term memory,LSTM)作为编码器-解码器模型的基本单元以提高模型捕捉长期依赖关系的能力,其次引入注意力机制调节编码向量的权重以进一步提高模型的预测精度。新的模型是一种典型的序列到序列预测模型,与传统的序列到点的模型相比更加契合日交通流预测的需求。为验证模型的有效性,取美国5号州际公路西雅图段的实际交通流数据进行实验,实验结果表明,提出的预测模型在平均车流密度大于40?辆/km的时间段中,其预测结果的平均绝对百分比误差(mean absolute percentage error,MAPE)与LSTM、门控循环单元(gated recurrent unit,GRU)、反向传播(back propagation,BP)神经网络、卷积神经网络(convolutional neural network,CNN)、图卷积网络(graph convolution network,GCN)传统预测模型相比,分别减小了19%、20%、25%、16%、25%。

关键词: 日交通流预测, 编码器-解码器, 深度学习, 长短时记忆网络(LSTM), 注意力机制

Abstract: Accurate daily traffic flow prediction is one of the important research contents in the field of intelligent transportation. At present, most of the existing daily traffic flow prediction models are improved by multi-step prediction or multi-objective prediction based on the short-term traffic flow prediction model. In these two improved schemes, the former is more sensitive to error propagation, and the latter ignores the time series relationship of prediction results, resulting in the low accuracy of the existing prediction models. This paper proposes an encoder-decoder deep learning model for daily traffic flow prediction. Firstly, long short-term memory(LSTM) neural network is taken as the basic unit of encoder-decoder model to improve the ability of the network to capture long-term dependence, and then attention mechanism is introduced to adjust the weight of coding vector to further improve the prediction accuracy of the model. Compared with the traditional sequence-to-point model, the new proposed model is a typical sequence-to-sequence model and is more suitable for the daily traffic flow prediction. In order to verify the effectiveness of the model, the actual traffic flow data on Interstate-5 in Seattle are used for the experiment. The experimental results show that the mean absolute percentage error(MAPE) of the prediction results of the proposed model is reduced by 19%, 20%, 25%, 16% and 25% respectively compared with LSTM, gated recurrent unit(GRU), back propagation(BP) neural networks, convolutional neural network(CNN) and graph convolution network(GCN) when the average traffic density is greater than 40 veh/km.

Key words: daily traffic flow prediction, encoder-decoder, deep learning, long short-term memory(LSTM), attention mechanism