短时交通流量预测是指15分钟内的交通流量预测，能为智能交通系统提供实时交通状况预测信息。现有短时交通流量预测模型在复杂交通环境中的预测误差仍高于10%。准确的数据源和高性能的预测模型是提高短时交通流量预测精度的关键，因此，本文以毫米波雷达为数据源，提出一种基于混沌理论和注意力机制的卷积长短时记忆网络，以实现短时交通流量预测。高测距、测速、测角以及稳定性的毫米波雷达能够满足模型对数据稳定、高精度的要求。同时，提出的预测模型相较于传统基于卷积循环神经网络的短时交通流量预测模型，能够更有效地解决长期依赖问题，实现了多场景下的交通流量高精度预测。研究内容包括：

（1）提出了一种新颖的短时交通流量预测算法（ACLSTM）以解决现有方法无法充分提取交通流数据特征的问题。使用卷积神经网络和长短时记忆网络相结合的方式，解决了交通流数据时空特征提取不完善的问题；在循环层后添加注意力机制，改善了模型无法聚焦于输入数据中最具预测意义的特征的问题；提出使用不同的周期输入矩阵，改善了交通流数据周期特性提取不全面的问题。经仿真验证，ACLSTM算法收敛速度快，预测误差比线性回归算法低了约23.29%，比CNN算法低了约15.9%，比GRU算法低了约14.4%。但ACLSTM算法在交通流数据非线性特征方面的建模能力不足，适用于资源受限的路况。

（2）使用混沌时间序列和混沌优化算法对ACLSTM算法进行改进，以解决ACLSTM算法在复杂场景下对交通流数据非线性特征捕捉不全面、模型泛化能力差的问题。提出使用相空间重构、递归图法和最大Lyapunov指数法相结合的方式判定交通流数据的混沌特性，验证了混沌理论对交通流数据的适用性；基于Lorenz系统构建了混沌时间序列，并将混沌时间序列作为训练模型的输入，改善了ACLSTM算法在交通流数据非线性特征方面建模能力不足的问题；设计了基于Logistic映射的混沌优化算法，并将其用于优化ACLSTM算法的各超参数，改善了ACLSTM算法泛化能力不足的问题。经仿真验证，Cts-ACLSTM算法在复杂场景下交通流量预测结果的平均绝对百分比误差（MAPE）为8.9%，而ACLSTM算法的MAPE指标为10.11%，Cts-ACLSTM算法的预测精度高于ACLSTM算法。

（3）短时交通流量预测及实测分析，对提出的Cts-ACLSTM算法的性能实施实景测试验证与对比。实验结果表明，所提算法的性能优越性与仿真结果一致，不同场景下1min时间间隔的流量预测误差均小于6.8%，15min时间间隔的流量预测误差均小于11%。与此同时，在不同时间间隔的交通流量预测中，周内预测误差和周末预测误差之间的波动均小于5%，实现了多场景下的交通流量高精度预测。

Short-term traffic flow prediction refers to predicting traffic flow within a 15-minute interval, which can provide real-time traffic condition prediction information for intelligent transportation systems. The prediction error of existing short-term traffic flow models in complex traffic environments still exceeds 10%. Accurate data sources and high-performance prediction models are crucial to improving the accuracy of short-term traffic flow prediction. Therefore, this paper proposes a convolutional long short-term memory network based on chaos theory and attention mechanism, with millimeter-wave radar as the data source, to achieve short-term traffic flow prediction. Millimeter-wave radar, with its high measurement range, measurement speed, measurement angle, and stability, can meet the requirements of the model for stable and highly accurate data. Moreover, compared with traditional short-term traffic flow prediction models based on convolutional recurrent neural networks, the proposed prediction model can more effectively solve the problem of long-term dependence, and achieve high-precision traffic flow prediction in multiple scenarios. The research content includes:

 

(1) Proposing a novel short-term traffic flow prediction algorithm (ACLSTM) to address the issue of existing methods not being able to fully extract traffic flow data features. The proposed method combines convolutional neural networks and LSTM networks, addressing the issue of incomplete extraction of spatiotemporal features in traffic flow data; self-attention mechanism is added after the recurrent layer to focus on the most predictive features in the input data; different periodic input matrices are utilized to resolve the incomplete extraction of periodic characteristics in traffic flow data. Simulation results show that the ACLSTM algorithm converges quickly, and its prediction error is approximately 23.29% lower than the linear regression algorithm, 15.9% lower than the CNN algorithm, and 14.4% lower than the GRU algorithm. However, the ACLSTM algorithm has inadequate modeling capabilities for nonlinear traffic flow data characteristics, making it suitable for resource-limited traffic conditions.

 

(2) Improving the ACLSTM algorithm using chaos time series and chaos optimization algorithms to address the issues of incomplete nonlinear feature capture and poor generalization capability in complex scenarios. The study proposes the use of phase space reconstruction and recurrence plot methods, combined with the maximum Lyapunov exponent method, to determine the chaotic characteristics of traffic flow data and verify the applicability of chaos theory to traffic flow data; using the Lorenz system to construct chaos time series and input them into the training model to address the inadequate modeling capabilities of the ACLSTM algorithm for nonlinear traffic flow data characteristics; using the Logistic mapping to design chaos optimization algorithms and optimize the hyperparameters of the ACLSTM algorithm, addressing the issue of insufficient generalization capabilities. Simulation results show that the Cts-ACLSTM algorithm has an average absolute percentage error (MAPE) of 8.9% in complex scenarios, while the ACLSTM algorithm has a MAPE of 10.11%, indicating that the Cts-ACLSTM algorithm outperforms the ACLSTM algorithm in prediction accuracy.

 

(3) The purpose of this study is to investigate short-term traffic flow prediction and to perform performance analysis of the proposed Cts-ACLSTM algorithm, followed by real-world testing validation and comparison. The experimental results demonstrate that the proposed algorithm has superior performance and the experimental results are consistent with the simulation results. In different scenarios, the proposed algorithm has a traffic flow prediction error of less than 6.8% in 1-minute time intervals, and less than 11% in 15-minute time intervals. Additionally, the fluctuation between the weekday prediction error and weekend prediction error in different time intervals of traffic flow prediction using this algorithm is less than 5%. This algorithm achieves high-precision traffic flow prediction in multiple scenarios.

[1] 刘洋, 路熙, 赵屾. 城市交通拥堵治理措施适应性研究 [J]. 交通运输研究, 2021, 7(05): 54-61.
[2] Hui Jie Yang. Xi 'an Intelligent Transportation System Construction Platform Research[J]. Pro-cedia Computer Science, 2019,154(C).
[3] 张艺,严翌瑄,李静. 基于多传感器融合的交通数据采集系统概述[J].物联网技术, 2021, 11(02):15-18.
[4] 杨颖,王立勇,孙鹏等. 基于激光雷达与相机融合的交通锥桶检测[J].激光杂志,2022,43(06):70-74.
[5] 王俊骅. 基于毫米波雷达组群的全域车辆轨迹检测技术方法. 中国公路学报35.12(2022):181-192.
[6] 李小路,周依尔,毕腾飞,余瑞钦,王子宁,黄建斌,徐立军.轻量型感知激光雷达关键技术发展综述[J].中国激光,2022,49(19):263-277.
[7] LUDENO G, CATAPANO I, RENGA A, et al. Assessment of a micro-UAV system for microwave tomography radar imaging [J]. Remote Sensing of Environment, 2018, 212: 90-102.
[8] SANCHEZ-ORO J, FERNANDEZ-LOPEZ D, CABIDO R, et al. Radar-based road-traffic moni-toring in urban environments [J]. Digital Signal Processing, 2012, 23(1): 364-74.
[9] WaveTronix. Patent Issued for Detecting Targets in Roadway Intersections[J]. Journal of Trans-portation, 2012.
[10] PENG Y, ABDEL-ATY M, SHI Q, et al. Assessing the impact of reduced visibility on traffic crash risk using microscopic data and surrogate safety measures [J]. Transportation Research Part C Emerging Technologies, 2017, 74(JAN): 295-305.
[11] 上海慧昌智能交通系统有限公司官方网站 [Z]. http://www.its-huichang.com.cn.
[12] 江苏志德华通信息技术有限公司官方网站 [Z]. http://www.t-radar.cn.
[13] 马莹莹, 靳雪振. 基于EEMD和小波阈值的短时交通流预测研究[J]. 重庆交通大学学报(自然科学版),2022,41(06):22-29.
[14] Ghaznavi-Youvalari R. Linear Model-Based Intra Prediction in VVC Test Model[C]// ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020.
[15] 张玺君,王晨辉.基于SARIMA-GA-Elman组合模型的短时交通流预测方法[J].兰州理工大学学报,2022,48(05):107-113.
[16] 王慧健,刘峥,李云,李涛.基于神经网络语言模型的时间序列趋势预测方法[J].计算机工程, 2019,45(07):13-19.
[17] 张兴锐,刘畅,陈哲,邓强强,吕明,罗谦.基于时空图卷积网络的机场地铁短时客流预测[J/OL].计算机工程与应用:1-9[2023-03-18].
[18] 黄益绍,韩磊.基于RS-IPSOSVM的公交客流量预测方法[J].重庆交通大学学报(自然科学版), 2020,39(11):11-19.
[19] Sun Z , Hu Y , Li W , et al. Prediction model for short-term traffic flow based on a K-means-gated recurrent unit combination[J]. IET intelligent transport systems, 2022(5):16.
[20] Bailin Y, Shulin S, Jianyuan L, et al. Traffic Flow Prediction Using LSTM with Feature En-hancement[J]. Neurocomputing, 2018,332.
[21] Markus Laner, Philipp Svoboda, Markus Rupp. Parsimonious Fitting of Long-Range Dependent Network Traffic Using ARMA Models.[J]. IEEE Communications Letters, 2013,17(12).
[22] Mai T, Ghosh B, Wilson S. Multivariate Short-Term Traffic Flow Forecasting Using Bayesian Vector Autoregressive Moving Average Model[C].Transportation Research Board 91st Annual Meeting, 2012,1:431-437.
[23] Okutani I, Stephanedes Y J . Dynamic prediction of traffic volume through Kalman filtering theory[J]. Transportation Research Part B, 1984, 18(1):1-11.
[24] 杨兆升,朱中.基于卡尔曼滤波理论的交通流量实时预测模型[J].中国公路学报, 1999, 12(3):63-67.
[25] 郭海锋,方良君,俞立.基于模糊卡尔曼滤波的短时交通流量预测方法[J].浙江工业大学学报, 2013,41(2):218-221.
[26] 贺国光,马寿峰,李宇.基于小波分解与重构的时间序列预测法[J].自动化学报,2002,28(6): 1012-1014.
[27] 金玉婷,余立建.基于小波神经网络的短时交通流预测[J].交通科技与经济,2014, 16(1):82-86.
[28] 王川,张宝文.基于小波分析与隐马尔科夫模型的短时交通流预测[J].交通节能与环保,2018, 14(1):43-47.
[29] Chand Sai. Modeling Predictability of Traffic Counts at Signalised Intersections Using Hurst Exponent[J]. Entropy, 2021,23(2).
[30] 承向军,刘军,马敏书.基于分形理论的短时交通流预测算法[J].交通运输系统工程与信息,2010,10(04):106-110.
[31] 张勇.交通流的非线性分形、预测和控制[D].[博士学位论文].北京:北京交通大学,2011:4-9.
[32] 王英平,王殿海,杨少辉,等.突变理论在交通流分析理论中应用综述[J].交通运输系统工程与信息,2005,5(6):68-71.
[33] 李洪萍,裴玉龙.基于混沌理论的交通流短时预测模型[J].昆明理工大学学报(理工版),2006, 31(5):95-99.
[34] Liu Y, Zhang J Z. Predicting traffic flow in local area networks by the largest Lyapunov expo-nent[J]. Entropy, 2016,18(1):1-10.
[35] 张玉梅,曲仕茹,温凯歌.基于混沌和RBF神经网络的短时交通流量预测[J].系统工程, 2007 (11):26-30.
[36] 杨庆芳,张彪,高鹏.基于改进动态递归神经网络的交通量短时预测方法[J].吉林大学学报(工学版),2012,42(04):887-891.
[37] 蒋士正,许榕,陈启美.基于变量选择神经网络模型的复杂路网短时交通流预测[J].上海交通大学学报,2015,49(02):281-286.
[38] Li Y F, Jiang X, Zhu H, et al. Multiple measures-based chaotic time series for traffic flow predic-tion based on Bayesian theory[J]. Nonlinear Dynamics, 2016,85(1):179-194.
[39] Vanajakshi L, Rilett L R. A comparison of the performance of artificial neural networks and sup-port vector machines for the prediction of traffic speed[C]. In Proceedings of the Intelligent Ve-hicles Symposium, 2004:582-588.
[40] 刘建华.相空间重构和SVR联合优化的短时交通流预测[J].计算机工程与应用,2014,50(3):13-17.
[41] 宫晓燕,汤淑明.基于非参数回归的短时交通流量预测与事件检测综合算法[J].中国公路学报, 2003, 16(1):82-86.
[42] 高勇,陈锋.基于小波分析的短时交通流非参数回归预测[J].中国科学技术大学学报,2008,38 (12):1427-1431.
[43] 张晓利,陆化普.非参数回归方法在短时交通流预测中的应用[J].清华大学学报(自然科学版),2009,49(9):1471-1475.
[44] Lin L, Guohong W, Xiangyu Z, et al. The track initiation algorithm based on Hough transform and space accumulation[C]//IEEE Beijing Section,IET Beijing Local Network,Beijing Jiaotong University.2016 IEEE 13th International Conference on Signal Processing Proceedings（ICSP2016）. Institute of Electrical and Electronics Engineers, 2016:1500-1504.
[45] Xu W , Hu J , Huang P , et al. Processing of Multichannel Sliding Spotlight SAR Data with Large Pulse Bandwidth and Azimuth Steering Angle[J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, PP(99):1-14.
[46] Jeong, Y.-S., Byon, Y.-J., Castro-Neto. Supervised Weighting-Online Learning Algorithm for Short-Term Traffic Flow Prediction[J]. IEEE transactions on intelligent transportation systems, 2013,14(4).
[47] Khandu Om, Spyros Boukoros, Anupiya Nugaliyadde, et al. Modelling email traffic workloads with RNN and LSTM models[J]. Human-centric Computing and Information Sciences, 2020,10(1).
[48] Lv T, Wang Y, Yuan J, Wu Y. Prediction and Application of Network Business Traffic based on LSTM[J]. Journal of Physics: Conference Series, 2022,2289(1).
[49] Zhou K, Zhan Y, Fu D. Learning Region-Based Attention Network for Traffic Sign Recogni-tion[J]. Sensors, 2021,21(3).
[50] Gu W, Tandon A, Ahn Yong Yeol. Principled approach to the selection of the embedding dimen-sion of networks[J]. Nature Communications, 2021,12(1).
[51] Small C M . Optimal embedding parameters: A modeling[J]. physica d nonlinear phenomena, 2003.
[52] Junhong F, Jie Z, Xiaoshu Z, Wenwu L. A novel chaos optimization algorithm[J]. Multimedia Tools and Applications, 2017,76(16).