申请同济大学工学硕士学位论文


改进的非参数回归方法

在短时交通流量预测中的应用研究

 

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:任千里

指导教师:石小法 副教授

二○一二年三月

 

A dissertation submitted to

Tongji University in conformity with the requirements for

the degree of Master of Engineering

The Advanced Nonparametric Model for

Short-Term Traffic Volume Forecasting

 

School: School of Transportation Engineering

Discipline: Transportation Engineering

Major: Transport Planning and Management

Candidate: Ren Qianli

Supervisor: Associate Prof. Shi Xiaofa

March, 2012

 

摘  要

智能交通技术已成为公认的解决日益严重的交通问题的最佳途径之一, 短期交通流分析及预测是智能交通系统的核心内容之一,也是先进的交通管理系统(ATMS)和先进的出行者信息系统(ATIS)的重要组成部分。因此,研究有效的短期交通流分析和预测的理论与方法,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之一。

本文围绕“如何建立一个面向城市快速路网实时应用的短时交通流量预测模型”展开研究,利用数据挖掘工具,以上海市快速路为案例进行了实证研究,论文主要完成了以下工作:

(1)对国内外已有的交通流量采集技术、短时交通流量预测的应用状况和相关预测理论进行了详细介绍; 在此基础上提出了改进的非参数回归方法预测短时交通流量的意义和可行性。

(2)结合上海市南北高架快速路的基础数据,对定点检测器采集的数据进行了分析和说明;针对原始数据中的错误、缺失等问题进行了数据预处理;最终将数据进行时空集成,得到了在一分钟、五分钟、十五分钟三个典型时间尺度下的一个月工作日的交通流量数据库,并进行了波动性及相关性分析。

(3)介绍小波神经网络(WNN) 、支持向量机回归(SVR)、非参数回归方法(NPR)的基本原理和本文采用的模型形式及相关参数;分别采用三种方法在三个时间尺度下进行了单点、单步的短时交通流量预测并对其预测效果、适用条件、影响因素等进行了全面的分析和评价。同等条件下支持向量机在 1 分钟、5分钟下预测精度最高,小波神经网络 15 分钟预测效果最好。同一时间尺度下基本路段的预测精度最高,其次是下匝道路段、交织区路段,上匝道路段的预测效果最差,应加强对快速路出入口尤其是上匝道对交通流量预测精度影响的研究。

(4)提出改进的非参数回归方法:在基本的非参数回归方法基础上进行了两个方面的改进:一、状态向量的选择和相似机制的改进——通过在上匝道处加入对支路流量的分析有效的提高了原有模型的预测精度,将平均相对误差从16.35%降低至 13.81%。二、历史数据库的精简——提出了基于滑动窗口控制的历史数据库精简模型(MW-NPR 控制模型),该模型在保证预测精度的前提下实现了历史数据的自组织和融合。

关键词: 短时交通流量预测,智能算法,改进的非参数回归模型,支路流量, 滑动窗口控制

 

ABSTRACT

As intelligent transportation systems (ITS) are implemented widely throughout the world, it has been proved as one of the most effective methods to ease traffic congestion, decrease traffic accident numbers and solve many other problems. Short-term traffic flow forecasting (usually less than one hour into the future) plays a key role in both the advanced traffic management systems (ATMS) and the advanced traffic information systems (ATIS). Therefore, the study of methods for short-term traffic flow analysis is not only the urgent need in the current development of ITS, but also one of the most important and difficult problems.

This paper focuses on real-time volume forecasting, which belongs to the short-term forecasting category, and the interest is on producing real-time forecast right after the system receives current flow data from an on-road traffic facility. Specifically, the emphasis of this paper is creating an accurately and efficiently database to predict the traffic volume in urban expressway of Shanghai. There are four parts in the dissertation.

(1) In the beginning we give a detail description of traffic flow adoption techniques and short-term traffic flow theory. Based on this we adapt the meaning and the practicability of the advanced nonparametric regression.

(2) The experimental data used for short-term traffic volume forecasting were acquired from the loop detector data from the eastern of Shanghai North-South urban expressway. For some issues of the original data such as errors, losses, noises and other, this paper studied the preprocessing techniques including data cleaning, smoothing, repairing and data generation. Finally, we calculate the data into different time steps of 1-min, 5-min, and 15-min respectively and analysis the fluctuation coefficient and similarity coefficient.

(3) To assess the accuracy of intelligent algorithm in traffic flow prediction under different time-steps, we introduce historical average as reference, wavelet neural network (WNN), support vector machine regression (SVR), nonparametric regression (NPR) as three typical intelligent methods to predict the traffic flow in single point of urban expressway in Shanghai. We calculate the mean absolute percentage error of four methods within 1-min, 5-min, and 15-min respectively and come to the following conclusions: SVR model obtains the best effect in 1-min and 5-min time-step. WNN is the best model when forecasting time is 15-min. Meanwhile, NPR model gets acceptable accuracy under three situations. Finally we discuss the impact of fluctuation coefficient and the road alignment to prediction accuracy. We also make recommendations to the issues and the future research direction.

 (4) Although the basic K-NN based nonparametric regression has been used in short-term traffic volume forecasting for a long time, some questions still exist. Firstly, state space selection and distance metric improved. According to the analysis of ramp volume, we improve the accuracy of original model. Secondly, this paper uses a new method called MW model to solve the problem that the database is too large and hard to search. Results show that the model was verified effective to achieve the data’s organization and merging itself with acceptable prediction accuracy.

Key Words:  short-term traffic flow prediction, intelligent algorithms, the advanced nonparametric regression, ramp volume, moving-window control model