交通运输系统工程与信息

交通运输系统工程与信息 ›› 2024, Vol. 24 ›› Issue (1): 115-123.DOI: 10.16097/j.cnki.1009-6744.2024.01.011

• 智能交通系统与信息技术 • 上一篇    下一篇

面向动态交通分配的交通需求深度学习预测方法

李岩1，王泰州1，徐金华1，陈姜会1，汪帆*1, 2   

  1. 1. 长安大学，运输工程学院，西安 710064；2. 中交第一公路勘察设计研究院有限公司，西安 710075
  • 收稿日期:2023-08-25 修回日期:2023-10-01 接受日期:2023-10-11 出版日期:2024-02-25 发布日期:2024-02-12
  • 作者简介:李岩(1983- )，男，河北衡水人，教授，博士
  • 基金资助:
    国家自然科学基金(51408049)；陕西省自然科学基础研究计划项目(2020JM-237)

Traffic Demand Prediction Method Based on Deep Learning for Dynamic Traffic Assignment

LI Yan1, WANG Taizhou1, XU Jinhua1, CHEN Jianghui1, WANG Fan*1,2   

  1. 1. College of Transportation Engineering, Chang'an University, Xi'an 710064, China; 2. China Communications Construction Company First Highway Consultants Co. LTD, Xi'an 710075, China
  • Received:2023-08-25 Revised:2023-10-01 Accepted:2023-10-11 Online:2024-02-25 Published:2024-02-12
  • Supported by:
     National Natural Science Foundation of China (51408049)；Natural Science Basic Research Plan in Shaanxi Province (2020JM-237)

PDF

摘要： 为满足动态交通分配对高精度、高时效性交通需求的要求，本文建立了一种交通需求深度学习预测方法。根据动态交通分配要求确定交通需求数据的时间间隔，构建对复杂交通需求预测性能较优的长短期记忆神经网络预测方法；针对动态交通分配中交通需求的周期性、随机性和非线性等特征，为减少数据噪声的干扰，引入局部加权回归周期趋势分解方法将交通需求数据分 解，将其中的趋势分量和余项分量作为深度学习预测方法的输入量，周期分量采用周期估计进行预测；选用具有随机寻优能力强、寻优效率高等特点的布谷鸟寻优算法优化预测方法的隐藏层单元数量、学习速率和训练迭代次数等核心参数。应用西安市长安区的卡口车牌数据验证该方法。结果表明：本文模型的预测结果在高峰及平峰各连续4个时段内相比于自回归滑动平均模 型、长短期记忆神经网络模型、支持向量回归模型，平均绝对误差降低了10.55%~19.80%，均方根误差降低了11.20%~17.99%，决定系数提升了8.62%~12.48%；相比遗传算法、粒子群算法优化的模型，平均绝对误差降低了7.36%~13.81%，均方根误差降低了4.23%~10.67%，决定系数提升了 3.50%~7.01%，且本文模型运行时间最短。说明与对比模型相比，本文所建立的预测方法在面向动态交通分配的交通需求预测中具有更高的预测精度。

关键词: 智能交通, 交通需求预测, 布谷鸟寻优算法, 长短期记忆神经网络, 动态交通分配, 局部加权回归周期趋势分解

Abstract: This paper proposes a deep learning traffic demand prediction method to meet the requirements of high accuracy and time sensitivity in dynamic traffic assignment. The time interval of traffic demand data is determined based on the requirements of dynamic traffic assignment. A prediction method using long short-term memory neural network is established for better performance in complex traffic demand. Combining the periodicity, randomness and nonlinearity of traffic demand in dynamic traffic assignment, this study uses a time series decomposition method to decompose the traffic demand data and to reduce the interference of data noise. The trend component and residual component are used as the input of the deep learning prediction method. Meanwhile, the periodic component is predicted using the cycles. The key parameters of the prediction method, such as the number of hidden layer units, learning rate and training iterations, are optimized by using the cuckoo search algorithm, which is characterized by strong random optimization ability and high optimization efficiency. The proposed method is verified using the checkpoint data in Chang'an District of Xi'an, China. In each of the four consecutive periods of peak and off peak, the results of proposed method are compared with the auto regressive moving average model, the long short-term memory model, and the support vector regression model. The results indicate a reduction of the average absolute error of 10.55% to 19.80%, a reduction of the root mean square error of 11.20% to 17.99%, and the coefficient of determination increased by 8.62% to12.48% . Compared with the models optimized by genetic algorithm and particle swarm optimization, the proposed model reduced the average absolute error by 7.36% to 13.81% and reduced the root mean square error by 4.23% to 10.67%. The coefficient of determination increased by 3.50% to 7.01%. The proposed model has the shortest running time. Compared with the traditional methods, the proposed prediction method has higher prediction accuracy in the traffic demand prediction for dynamic traffic assignment.

Key words: intelligent transportation, traffic demand prediction, Cuckoo search algorithm, long short-term memory; dynamic traffic assignment, seasonal and trend decomposition using loess

中图分类号: