申请同济大学工学博士学位论文


基于智能算法的交通流量及状态预测研究

 

培养单位:交通运输工程学院

一级学科:交通运输工程

二级学科:交通运输规划与管理

研 究 生:杨 文

指导教师:杨东援 教授

二○一○年十二月

 

A dissertation submitted to

Tongji University in conformity with the requirements for

the degree of Doctor of Philosophy

Traffic Flow and State Prediction Based on

Intelligent Algorithms

 

School: School of Transportation Engineering

Discipline: Transportation Engineering

Major: Transport Planning and Management

Candidate: Yang Wen

Supervisor: Prof. Yang Dong-yuan

December, 2010

 

摘  要

智能交通系统被认为是缓解道路交通拥堵、减少汽车尾气排放污染和交通事故等交通问题的有效方法之一。短期交通流分析是智能交通系统的核心内容,交通流的准确预测和状态判别是交通信息服务、交通控制与诱导的重要基础。因此,研究有效的短期交通流分析和预测的理论与方法,从所获得的信息中准确地预测和判别出道路交通状态,是当前智能交通系统发展的亟需,也是研究的重点和难点问题。

本文要解决的问题就是如何从带有随机性和不确定性的交通流变化中,进行交通流参数的分析,找出其中的规律性,建立相应的预测方法和模型并进行改进,提高预测精度,在此基础上,对未来时段的交通流状态进行预测。

针对断面交通检测数据往往存在着错误、缺失、包含较多噪声等问题,研究了原始交通流数据的预处理技术,包括数据的清洗、平滑、修补及地点车道组数据的生成。

提出了相似系数和波动系数作为交通流周期相似性的判定依据。以上海内环高架的主线线圈为例,根据其交通流的周期相似性把线圈分为三类,分别进行流量预测,并分析比较预测结果,验证交通流的可预测性。

通过利用小波分析和RBF神经网络,建立了一种小波神经网络组合模型。该模型对周期性交通流量进行小波分解与重构后,用RBF神经网络分别对各层分量进行预测,最后对各层分量的预测值进行合成。在实际案例中,采用多种时间尺度对不同预测模型的预测结果进行比较。

由于在对交通流量进行预测时,选取的历史数据对预测结果影响较大,通过对每周交通流量相似系数和不同周同一工作日相似系数进行分析,运用最小二乘支持向量机回归进行交通流量的特征曲线提取,选取具有相似性的交通流量作为训练数据,以LSSVM作为预测模型,对未来5天的交通流量进行预测。

通过对混沌交通流的特性进行分析,确定重构相空间所需参数的计算方法及可预测时间尺度,提出一种改进的加权一阶局域预测模型,该模型同时考虑相点间的欧氏距离和相似性大小来确定邻近点权重。预测结果表明,改进的加权一阶局域预测法具有比一般加权一阶局域预测法更好的预测精度。

在用最小二乘支持向量机建立多步预测模型,选取训练样本时,提出一种改进的邻近点选取方法,即首先应用欧氏距离确定暂时邻近点,然后选取满足暂时邻近点与预测中心点间一定相似度的邻近点作为最邻近点,并用实际案例分别对一般方法和改进方法的预测效果进行比较分析。预测结果显示,将相空间重构理论和最小二乘支持向量机结合起来所建立的预测模型能够捕捉到原混沌交通流的动力学特征,采用改进的邻近点选取方法进行预测具有更好的预测精度。

设计了基于固定型交通检测数据的快速路交通流状态预测方法。选取流量和速度作为评价交通流运行状况的指标,并分别制定各指标的模糊集和用以判别交通状态的模糊规则。以上海市南北高架上的19个检测线圈为例进行具体分析,预测快速路交通流的动态演变趋势,并对研究路段分别用畅通、稳定、拥挤和堵塞四类交通流状态进行状态划分。

关键词: 流量预测,相似系数,小波分析,RBF神经网络,相空间重构,最小二乘支持向量机,模糊逻辑

 

ABSTRACT

ITS(Intelligent Transportation System) is considered as one of the most effective methods to ease traffic congestion , reduce vehicle emissions , decrease traffic accident numbers and solve many other traffic problems.Short-term traffic flow analysis is the core of ITS. Exact prediction and state estimation of traffic flow are the foundation of traffic information service , control and guidance. Therefore, the study of effective theories and methods for short-term traffic flow analysis and prediction, to accurately predict and determine the road traffic state with the information obtained, is not only the urgent need in the current development of ITS in practice,but also one of the most important and difficult problems.

This paper will focus on the analysis of traffic flow parameters to find out the regularity from the randomness and uncertainty of the traffic flow changes,the establishment of appropriate methods and models to improve the prediction accuracy.On this basis, traffic flow state of future periods will be predicted.

For these issues of cross section detection data,such as errors, losses, noises and others,this paper studied the preprocessing techniques of original traffic flow data,including data cleaning, smoothing, repairing and data generation of lane group at the same place.

The similarity coefficient and the fluctuation coefficient were put forward as the determining criteria of traffic flow cycle similarity. Main coils of the Inner Ring Elevated in Shanghai were selected as analysis examples.These coils were then divided into three categories and predicted respectively according to the cycle similarity of traffic flow. Furthermore , the predicted results were analyzed and compared to verify the predictability of traffic flow.

This paper established a kind of wavelet neural network model combing with the advantages of both wavelet transform and RBF network.Traffic flow data with similar periods were selected and made a wavelet decomposition and reconstruction for flow-time series, then signal components were predicted by RBF neural network,and prediction results were then synthesized. In the actual case, the prediction results using a variety of time scales were compared with different prediction models.

The selection of historical data is essential to predict the short-term traffic flow, different historical data may result in quite diversified prediction results.The traffic flow similarity of each week and that of the same working day in different weeks were analyzed respectively. The characteristic curve of traffic flow was extracted using least squares support vector machine regression,and then the similar traffic flow data were selected as the training data. LSSVM was used for the prediction model,and continuous 5-day prediction was carried out.

Based on analyzing the characteristics of chaotic traffic flow ,this paper established the calculating method of required parameters for phase space reconstruction.Then an improved adding weight one-rank local region prediction model was proposed.The model took into account the weights of Euclidean distance and similarity between phase points.Prediction results of an actual case show that the prediction accuracy of improved model is better than that of the general model.

An improved method for the selection of training samples was proposed,and the multi-step prediction model was established with least squares support vector machines. After the selection of temporary neighboring points using the Euclidean distance, these temporary neighboring points must satisfy certain similarity coefficient. In the actual case,the prediction results of the general method and the improved method were compared.It can be concluded that the prediction model combined with phase space reconstruction and LSSVM could catch the dynamics characteristics of chaotic traffic flow, and the prediction accuracy is much better after the selection of improved method.

An prediction method of traffic state was designed according to the fixed detection data of expressway.Volume and speed were selected as the indicators to evaluate the operating status of traffic flow,and the fuzzy sets of the indicators and the fuzzy rules for the determination of traffic state were designed respectively. 19 detection coils of North-South Elevated Road in Shanghai were selected as analysis examples.The dynamic evolution trend of traffic flow was predicted,and the traffic state of the object link was classified as follows: free, stability, crowd, congestion.

Key Words:  flow prediction, similarity coefficient, wavelet analysis, RBF neural network,phase space reconstruction, least squares support vector machine, fuzzy logic