【作者】 王凡；

【导师】 谭国真；

【作者基本信息】 大连理工大学 ， 计算机软件与理论， 2010， 博士

【摘要】 实时准确的交通流量预测是实现智能交通控制和诱导的前提与关键,也是智能化交通管理的客观需要。由于城市交通系统本身的强非线性、随机性、时变性等特点,以精确数学模型为基础的传统预测方法效果并不理想,因此人工智能的方法愈来愈受到人们的重视。支持向量机(Support Vector Machine, SVM)作为一种基于统计学习理论的机器学习方法,较好地解决了非线性、高维数、局部极小点等实际问题,是复杂非线性科学和人工智能科学的研究前沿。本文总结了交通流预测模型的研究现状,比较了预测模型的优缺点,分析了交通流数据的相关性和影响交通流变化的主要因素,以此为基础,主要从支持向量机回归方法进行研究和其在交通流预测中的应用相结合的角度出发,对支持向量机的参数选择、核函数构造、增量学习、并行计算等方法以及在交通流预测中的应用进行了系统的研究,主要工作如下：(1)对支持向量机参数选择进行了研究。不敏感损失系数、惩罚系数、核函数及其参数的优化选择对回归模型的学习精度和推广能力的好坏起着重要的作用。结合交通流数据特性,本文提出自适应参数选择的交通流预测方法,利用训练集对支持向量回归参数进行求解。与BP神经网络方法、经验选择方法和粒子群优化方法进行比较,自适应参数选择方法能够依据训练样本集自适应进行参数选取,实现预测模型自适应能力,有效地改善了交通流量的预测精度。(2)对支持向量机核函数的构造方法进行研究。目前较为常用的核函数为高斯径向基核函数,它在模式识别与回归分析中都表现出了良好的映射性能,但是径向基不是一个完备基,不能较好地对任意信号逼近,特别是对边界处逼近和多尺度信号逼近。本文利用子波良好的时频多分辨特性及逼近特性,构造基于Marr子波核函数和多尺度核函数,通过仿真试验分析,在泛化性能和训练时间方面优于传统径向基核函数,同时弥补了传统核函数在逼近性能方面的不足,较好地解决了交通流量实时预测中存在的随机干扰因素影响大、不确定性强的问题。(3)对支持向量机并行计算技术进行研究。对于大规模路网的动态交通流预测,并行计算是解决短时交通流预测中对大样本集进行快速训练的一条重要途径。本文在研究并行SVM技术的基础上,引入并行SMO算法用于交通流预测,同时对2000条多路段并行实验研究,通过在深腾1800高性能机器上进行实验,验证了基于SVR大规模交通流预测的实用性和可用性。更多还原

【Abstract】 Accurate real time traffic flow prediction is a prerequisite and key to realize intelligent traffic control and guidance, and it is also the objective requirement to intelligent traffic management. Due to the strong non-linear, stochastic, time-varying characteristics of urban transport system, traditional forecasting methods based on accurate mathematical models is not ideal, therefore, artificial intelligence methods obtain more and more attention. Support Vector Machine (SVM) is a machine learning method based on statistical learning theory (SLT), it can efficiently solve small samples, nonlinear, high dimensions and local minima problems, and it is a research front of complex non-linear science and artificial intelligence scientific research. In this paper, the research situation of traffic flow prediction models are summarized, the advantages and disadvantages of different models is compared. Based on the analysis of the correlation of traffic flow data and important factors which affect the traffic flow’s change, the support vector regression method and its application in traffic flow prediction are systematically discussed, such as the support vector machine’s parameter selection method, kernel function construction method, incremental learning technique, parallel computing technology and its application in traffic flow prediction, main tasks are as follows:(1) Research on the SVM parameter selection. The selection of insensitive loss coefficient, penalty coefficient C, kernel function and its parameters is important to the learning accuracy and generalization ability of regression model. In this paper, according to the traffic data properties, a traffic flow prediction method with adaptive parameter selection is proposed, using the training set to solve support vector regression parameters. Compared with traditional experience selection SVR method, adaptive parameter selection SVR method can adaptively select parameters based on training set, realize the model’s adaptive capacity and efficaciously improve traffic flow prediction accuracy.(2) Research on support vector machine kernel function’s construction method. Since the widely used Gaussian (Radial Basis Function) kernel function can not does arbitrary signals approximation well, especially the boundary approximation and multi-scale signals approximation, the wavelet is used which has good multi-resolution time-frequency characteristics and approximation properties, and SVR with Marr wavelet kernel function and multi-scale kernel function are constructed. With simulation test analysis, this method has better generalization capability and training speed, make up the approximation performance deficiencies of traditional kernel functions and solve the stochastic interference factors’ enormous influence and strong uncertainty problems in real time traffic flow prediction.(3) Research on support vector machine’s parallel computing technology. In the dynamic traffic flow prediction in large-scale network, parallel computing is an important solution to solve the large sample set’s fast training problems in short time traffic flow prediction. With the research on parallel SVM, parallel SMO is employed in traffic flow prediction. Simultaneously,2000 sections experiments are carried on the DeepComp 1800 high-performance machine. The experimental results demonstrate parallel SVR is practicable and effective to predict traffic flow in large-scale traffic network.更多还原