张伟涛

入学时间：2019级

答辩时间：2022年

论文题目：基于智能计算的信号控制交叉口交通状态识别与预测研究

中文摘要

摘要

城市道路交通运行状态识别和预测对于其最佳运行管理与控制具有支撑作用。信号控制交叉口作为道路网络的节点，其运行状态感知与预测更为重要。交叉口交通拥堵不仅会直接关联其上下游路段，甚至可能影响整个路网。因此，研究交叉口交通状态识别与预测方法，可以为其交通管理与控制提供决策依据，进而可有效地缓解交叉口交通拥堵、降低排放与能耗，并提高其通行效率。

关于交叉口交通状态的定义和量化，目前存在多样标准，因此其表征指标也有不同，指标选取存在一定的主观性，其合理性、科学性与有效性尚需要进一步验证。现有研究多以实际经验确定交通状态分级取值，但这种方法通常具有主观性和随机性，与交叉口实际情况的适应度不高。并且，交通状态识别和预测研究多基于交通流的时间序列，提出多种模型，并在高速公路、快速路等场景中应用较好。然而，当场景聚焦于信号控制交叉口时，由于信号控制作用、通行环境的复杂、车流频繁变化与交汇、交通需求的不确定性等因素影响，其交通状态呈现出较强的波动性和随机性特点，以往基于模型计算的交通状态不能充分地体现交叉口交通状态的变化，也不能有效地改进随机因素对识别及预测产生的影响，存在鲁棒性不高、算法精度较低、应用效果差等问题。

智能计算的应用为上述研究带来了新机遇，其方法普遍具有自学习与自适应等特征，能够助推交叉口交通状态及其表征参数关系更加明晰，进而可以更加有效地识别和预测交通状态。本文聚焦于信号控制交叉口场景，在总结信号控制交叉口理论研究现状基础上，进行了交通参数表征问题分析研究；建立了交通状态划分的模拟退火-遗传算法的模糊C均值聚类（SAGA-FCM）的算法，交通状态判别的分层重采样的随机森林（SR-RF）算法，交通状态预测的误差补偿方法与思维进化算法结合的小波神经网络（EC-MEA-WNN）算法。详细如下：

(1)针对交叉口交通状态基础理论问题，总结了常见状态指标及交通参数，分析了状态影响因素，选取饱和度、延误、平均排队长度组成的交叉口交通状态表征体系，通过视频得出的实际数据结合排队论，分析了交通流在交叉口的相似性、交叉口的排队现象与阻滞现象，论证了本文选取的交通状态表征体系的合理性。

(2)针对交叉口交通状态划分问题，提出了采取模糊C均值聚类算法划分交通状态，以聚类有效性函数决定分类数目，并将交通状态划分为四类的方法，且针对原算法初始聚类中心为随机生成，容易陷入到局部最优的问题，采用模拟退火-遗传算法加以改进。实证分析表明，SAGA-FCM算法的收敛速度约为FCM算法的2倍，且多次运行算法后，SAGA-FCM算法的目标函数值保持为6623.14不变，较FCM算法的稳定性更好，证明了本文所提出的算法在收敛速度和稳定性上的优势。

(3)针对交叉口交通状态判别问题，研究提出了随机森林方法，并为提高模型对不平衡数据集的判别精度，使用分层重采样方法改进训练集。实证表明，本文提出的基于分层重采样的随机森林方法，对少数类样本的精确率和召回率较改进前分别从86%提升至88%和100%，且综合评价指标F1值（F1-Measure）从0.88提升至0.93；对比BP神经网络模型，本文方法的判别准确率较之提高了10.5%，F1值提升了8.5%，证明了本文方法的可行性。

针对交叉口交通状态预测问题，本文采用小波神经网络方法对交叉口交通状态表征参数进行预测，结合状态判别方法实现交通状态间接预测。有关预测误差中隐含规律和单一小波神经网络预测模型中的梯度下降法寻优过程中易陷入极值问题，采用误差补偿方法和思维进化算法进行优化。实证分析表明，本文提出方法在预测评价指标和状态预测结果上均有改善，其中，平均绝对百分比误差（MAPE）与均方根误差（RMSE）改善较为明显，MAPE平均减小38%，RMSE平均减小13%；未能正确预测的交通状态样本数从13降低至7，证明了本文所提出方法的有效性。

关键词：信号控制交叉口，交通状态，智能计算，状态识别，状态预测

英文摘要

ABSTRACT

In the urban road network, the identification and prediction of traffic operation status play a supporting role in its optimal operation management and control. The signalized intersection is the node of the road network, the perception and prediction of its operation are more important. The congestion of the intersection will affect its upstream and downstream sections, and even affect the entire road network in severe cases. Research on the method of identification and prediction can provide a decision-making basis for traffic management and control, it can effectively ease traffic congestion at the intersection, reduce emissions and energy consumption and improve the traffic efficiency of the intersection.

There are many standards for the definition and quantification of the traffic status at the intersection, the selected status representation indicators are different. The selection of indicators is subjective, the rationality, scientificity, and effectiveness of the indicators need to be further verified. Existing research mostly determines the number of traffic statuses based on experience, but this method is usually subjective and random, and the adaptability to the actual situation of the intersection is not high. In addition, many traffic status identification and prediction studies have proposed a variety of models based on the time series of traffic flow, which are well applied in scenarios such as highways and expressways. However, in signalized intersections, due to the volatility and randomness caused by factors such as signal control, complex traffic environment, traffic conflict, and uncertainty of traffic demand, the above models cannot well reflect the traffic status of the intersection, nor can they handle the influence of random factors on identification and prediction, and there are problems such as low robustness, low algorithm accuracy, and poor application effect.

The application of intelligent computing has brought new opportunities for the above research. Intelligent computing methods generally have the characteristics of self-learning and self-adaptation, which can help to make the relationship between the traffic status and the parameters of signalized intersections clearer, and traffic status can be more effectively identified and predicted. In this paper, taking the signalized intersection as the research object, based on summarizing the basic theory of signal-controlled intersection and the relevant research, it analyzes the characterization of traffic parameters and proposes a method for dividing the traffic status of the intersection based on the fuzzy C-means clustering based on simulated annealing-genetic algorithm (SAGA-FCM) algorithm. An intersection traffic status discrimination method based on random forest based on stratified resampling (SR-RF) algorithm is proposed, and an intersection traffic status prediction method based on wavelet neural network based on error compensation-mind evolution algorithm (EC-MEA-WNN) algorithm is proposed. The specific work is as follows:

(1) Aiming at the theoretical study of the traffic status of signalized intersections, the commonly used status evaluation indicators and the traffic parameters of the intersection are summarized, and the influencing factors of the traffic status of the intersection are analyzed. Then, saturation, delay, and average queuing length are selected to constitute the intersection traffic status representation system. By combining the measured data obtained from the video with the queuing theory, the similarity of the traffic flow at the intersection, the blocking phenomenon, and the queuing phenomenon of the intersection were analyzed, and the rationality of the selected parameters on the traffic status was analyzed.

(2) For the division of the traffic status of the intersection, this paper adopts the fuzzy C-means clustering method to divide traffic status represented by the traffic flow parameters, determines the number of traffic statuses by the clustering validity function, and divides the traffic status of the intersection into four types. The fuzzy C-means clustering algorithm has random initial clustering centers, it is easy to fall into local optimum, and the simulated annealing-genetic algorithm is used for optimization. Through empirical analysis, the convergence speed of the SAGA-FCM algorithm is about twice that of the FCM algorithm, after running the algorithm for many times, the objective function value of the SAGA-FCM algorithm remains unchanged at 6623.14, which is more stable than the FCM algorithm, the advantages of the algorithm proposed in this paper in terms of convergence speed and stability are proved.

(3) For the identification of the traffic status of the intersection, a traffic status identification model of the intersection based on the random forest is proposed. To improve the identification accuracy of the model for the imbalanced data set, the training set is modified by the stratified resampling method. Through example demonstration, the precision and recall of the random forest method based on stratified resampling proposed in this paper for minority class samples increased from 86% to 88% and 100% respectively, and the comprehensive evaluation index F1-Measure increased from 0.88 to 0.93. Compared with the BP neural network model, the accuracy of the method in this paper is increased by 10.5%, and the F1-Measure is increased by 8.5%. which proves the feasibility of the method proposed in this paper.

(4) For the prediction of the traffic status of the intersection, this paper uses the wavelet neural network method to predict the parameters representing the traffic status of the intersection, and combines the status identification method to realize the indirect prediction of the traffic status. Because of the implicit rule in the prediction error and the gradient descent method in a single WNN prediction model, it is easy to fall into the extreme value in the optimization process, and the error compensation method and the mind evolution algorithm are used for optimization. Through case analysis, the proposed method for predicting the traffic status of the intersection has improved both the prediction evaluation index and the status prediction result, the mean absolute percentage error (MAPE) and root mean square error (RMSE) improved significantly, with an average reduction of 38% in MAPE and 13% in RMSE, which proves the effectiveness of the method proposed in this paper.

Key Words: Signalized intersection, Traffic status, Intelligent computing, Status identification, Status prediction