【作者】 史慧敏；

【导师】 谭国真；

【作者基本信息】 大连理工大学 ， 计算机应用技术， 2009， 硕士

【摘要】 智能交通系统(Intelligent Transportation Systems,ITS)是目前国际公认的缓解城市交通拥挤的最佳途径。道路交通信息是所有ITS项目不可缺少的前提和主要内容。如何在短时间内得到这些信息,以及如何根据这些信息快速确定出最佳行驶路径,已成为ITS领域的一个前沿问题,交通路网短时交通信息预测理论、模型与算法的优劣直接影响整个ITS的实施。交通状况信息中最基本的参数就是通过路段的交通流量和旅行时间,它们代表着路段的物理属性和交通特性,也是用户所关心的最直接指标。短时交通信息预测可以通过城市交通信息发布平台为出行者提供实时的交通信息,帮助他们进行路径的选择及诱导。人工神经网络由于其良好的非线性映射能力,已在交通预测中得到广泛应用。训练神经网络常采用BP(Back Propagation)算法,但BP算法具有收敛速度慢,易陷入局部最小的缺点。为了加快神经网络的学习速率,许多并行学习算法被相继提出,本文采用一种并行非线性优化技术训练神经网络,实现有检测器路段的交通流预测。利用并行变尺度拟牛顿法(Self-Scaling Parallel Quasi-Newton,SSPQN)改进BP算法,每次迭代时产生多个搜索方向,各并行子任务在不同的方向上执行非精确线性搜索以寻找最优点。并在MPICH并行环境下对上述算法进行测试分析。实验结果表明在达到相同训练精度的前提下,SSPQN算法有效提高了收敛速度,预测效果优于BP算法,基本达到实测路况交通流预测的要求。为了实现路网中任意路段的流量预测,针对无检测器路段,使用多元统计分析中的多维标度法对城市路网相关性进行分析,实现对整个路网的划分,借助同类交叉口的有检测器路段的流量对无检测器路段的流量进行预测。最后采用旅行时间函数模型经过参数标定,建立适用于国内城市的旅行时间估计模型,进而建立短时旅行时间的预测模型,并在VISSIM仿真平台上对该模型进行验证,实验结果表明满足动态交通诱导的要求。更多还原

【Abstract】 Intelligent Transportation Systems (ITS) is the internationally recognized best way to relieve urban traffic congestion. Road traffic information is an indispensable prerequisite and main content for all ITS projects. The problem that how to get these messages in a short time and how to use these informations to determine quickly the optimal driving path has become a cutting-edge issues in the field of ITS, traffic network short-time traffic information prediction theory, the pros and cons of model and algorithm directly affect the whole ITS implementation. The traffic flow and travel time through section is the most basic parameters of traffic information. It not only represents the physical properties and transport properties of some section, but also is the most direct interest concerned by users. Short-term traffic information prediction can provide real-time traffic information through urban traffic information release platform for the trip, also help them to the path selection and induction.The artificial neural network is a new method of maths modeling. Because of better adaptability, it has been widely applied in traffic forecasting. Back-propagation (BP) algorithm is usually used to train the neural network, but it converges slowly and easily gets into local minimum. To speed up learning of neural network, many parallel learning algorithms are proposed. This thesis uses a kind of parallel nonlinear optimization technique to train networks, realizing parallelism based on learning algorithm. BP algorithm is improved by parallel self-scaling quasi-Newton (SSPQN) algorithm. In each iteration, a set of search directions is generated. Each subtask carries out inexact linear search along each direction to find the optimal point. This thesis tests and analyses the above algorithm in the MPICH parallel environment. The experimental result shows that with the same training precision the SSPQN parallel algorithm effectively improves the convergence speed, has better forecast effects than traditional BP neural network, and meets the requirement of traffic flow forecasting.In order to achieve a prediction of traffic volume in arbitrary section of the network, the thesis use multivariate statistical analysis of the multidimensional scaling of urban road network correlation analysis for non-detector section, it achieves the division of the entire network, and predictions the flow of the non-detector section by using the flow rate of the same intersection with a detector section. Finally, the thesis estimates the travel time using travel model time after the parameter calibration, meanwhile it establishs a short-term travel time prediction model, and validates the model in the VISSIM simulation platform. Experimental results show that it meets the dynamic traffic-induced demand.更多还原