杨超兰
入学时间：2015级
答辩时间：2018年
论文题目：信号控制交叉口交通安全指数计算方法研究
中文摘要
摘要
道路交通系统是一座城市的重要骨架和支撑，交叉口是道路网络的关键节点，各类交通流在此交汇，因此存在较大的交通事故风险。深入研究信号控制交叉口的交通安全问题，构建科学有效的安全评价方法与体系，对于揭示交通安全规律，提高交通系统安全性，乃至保障我国居民生命财产安全具有重要的理论价值和现实意义。本文在总结国内外信号控制交叉口交通安全评价相关研究的基础上，提出“信号控制交叉口交通安全指数”的概念，从交叉口事故致因分析，综合考虑各类影响因素，运用交通冲突和系统工程等理论方法，对信号控制交叉口交通安全指数的计算方法进行了研究。
首先，论文研究了信号控制交叉口交通安全影响因素及其作用。运用交通微观仿真（VISSIM）与交通安全替代分析软件（SSAM）相结合的方法，定量分析了交叉口饱和度、倒计时信号灯和相位相序对交叉口安全性的影响，得到了不同交叉口饱和度下的交通冲突数的变化特性；针对不同的道路特征，重点分析了交叉口渠化对交通安全性的影响，建立了右转车辆与行人之间的潜在交通冲突数估计模型；有关管理控制条件的影响，选取了绿灯间隔时间和行人过街最短时间作为信号控制方案中交通安全性的重要影响因素，并给出了不同交叉口几何特征与行驶速度下的最短全红时间和最短绿灯间隔时间的参考值。最后探讨了非机动车、道路特征、管理控制条件、行为特征和通行环境对交叉口安全性的影响。
其次，论文借鉴交通事故发生的4E过程，引入了交通暴露状况和交通冲突状况作为信号控制交叉口交通安全性的主要影响因素；在此基础上，构建了交通安全指数的计算思路，并对计算方法和参数取值进行了深入研究，给出了交通安全指数分级标准。然后，选取了上海市三处信号控制交叉口作为应用对象，采用交通安全指数的计算方法，得到了三处交叉口的交通安全指数及其安全等级评价，论文进一步分别采用交通仿真法和专家调查法，验证了交通安全指数计算方法的有效性及合理性。
最后，论文归纳了主要的研究成果和创新点，并对进一步的研究方向进行了简要讨论。
关键词：信号控制交叉口，交通安全指数，交通冲突，交通暴露，交通安全替代分析软件（SSAM）
英文摘要
ABSTRACT
The road traffic system is an important framework of a city, and intersections are the key node of road traffic system. Various types of traffic flow converge at intersections, increasing the risk of traffic accidents. So there are important theoretical values and realistic meanings to study the traffic safety evaluation method of signalized intersections to improving safety and protecting the lives and property of our residents. Based on some related research, the dissertation introduced Traffic Safety Index of signalized intersections considering the factors of accident and studied the calculation method of the index by applying knowledge of traffic conflict technologies and system engineering.
Firstly, the traffic safety influence factors at signalized intersections were studied. Combined traffic microscopic simulation software (VISSIM) with Safety Surrogate Analysis Model (SSAM), a quantitative analysis of effects of traffic saturation, countdown lights and phase sequence on the intersection safety was made, and the number of traffic conflict at different traffic saturations are obtained. In the road characteristics, the influence of channelization on traffic safety is analyzed, and potential traffic conflict estimation model between right-turning vehicles and pedestrians is established. In the management and control, change interval and the shortest time of pedestrian crossing are selected as the important factors of traffic safety in the intersection signalization scheme, and reference value of the shortest all-red time and the shortest change interval at different intersection geometric features and driving speed is given. Then, the impact of non-motorized vehicles, road characteristics, management and control conditions, human behavior and environmental factors on the safety of intersections was discussed.
Secondly, based on road safety 4E process, the dissertation introduced traffic exposure and traffic conflict as the main factors of Traffic Safety Index. The calculation method of Traffic Safety Index of signalized intersections was constructed and the classification standard was obtained. Thirdly, three signalized intersections in Shanghai were selected as applications. The Traffic Safety Index and safety level at these intersections were obtained by the calculation method proposed in the dissertation. Then, two methods of Traffic Simulation and Expert Investigation Method were used separately to analyze the traffic safety level at the three intersections. The results showed that the traffic safety index and other two safety evaluation results are consistent.
Finally, the dissertation is summarized, the creative research achievements and other important research directions were pointed out.
Key Words:signalized intersection, Traffic Safety Index, traffic conflict, traffic exposure, Safety Surrogate Analysis Model
刘洋东
入学时间：2015级
答辩时间：2018年
论文题目：基于深度学习的城市道路行程时间预测方法研究
中文摘要
摘要
行程时间预测是ITS技术和先进的交通管理系统（Advanced transportation management systems, ATMS）发展和应用的重要组成部分，对于管理者制定主动交通管理政策和出行者规划合理高效的出行路径具有重要意义，也是推动交通流理论发展的重要理论研究方向。然而，城市道路行程时间受到交叉口信号控制、交通需求的动态性和不确定性及大量交通事件的影响而呈现出剧烈波动性和强随机性特点，使准确、稳健预测的难度大大增加。同时，随着机器学习领域的理论革新、高性能计算和大数据技术的逐渐兴起，深度学习渐渐从无法实现的概念发展成为可以超越众多模型实现准确预测和进行大规模应用部署的复杂机器学习模型，因而受到越来越多的研究者关注。
既有的部分浅层学习预测方法具有脆弱性，浅层性和有限性等特点，对突发性交通事件难以准确建模，无法提取丰富特征，只对特定时间段内的预测有效，因而无法成功解决城市道路行程时间预测的问题；而深度学习方法提取特征层次丰富，能较好地对具有时间依赖的数据建模。然而深度学习也面临训练开销大，超参数优化困难的挑战。本研究旨在建立基于深度学习的城市道路行程时间预测模型与方法，并从稳定性、准确性等多个角度对比既有研究中有代表性的浅层学习模型与方法。本文主要研究内容如下：
1)深入对比分析既有浅层学习与深度学习方法的发展情况，并提出基于滑动时间窗的短时动态预测机制；
2)提出一种适用于城市道路行程时间预测的LSTM-DNN深度学习模型，并实现经典的浅层学习模型（如线性回归，岭回归，LASSO回归，随机森林等）以进行预测性能的全面综合对比；
3)为了构建最优的深度学习模型，本文创新性地提出一套基于网格搜索，多预测场景调优和多组训练法的深度学习超参数最优化方法。多组训练法可以规避深度学习模型随机性问题。基于2017年深圳市城市道路行程时间数据，构建10个不同预测场景进行了模型验证，以全面综合对比特定超参数下的模型性能。
4)建立了多种基于浅层学习和深度学习模型为次级学习器的集成学习模型，探究集成学习模型是否能够综合浅层学习和深度学习模型的不同特点以提升城市道路行程时间的预测水平。
关键词：行程时间预测，深度学习，循环神经网络，长-短期记忆网络（LSTM），超参数优化，集成学习
英文摘要
ABSTRACT
Travel time prediction is an important issue of the development and application of ITS techniques and Advanced Transportation Management Systems (ATMS). It is important for transportation managers to develop active traffic management policies , for traffic system users to plan reasonable and efficient travel routes, and for the development of theoretical research on traffic flow theory. However, travel time on urban road segments is characterized by drastic fluctuations and randomness due to signal control at intersections, dynamic and uncertainty of traffic demand and a large number of traffic incidents, which greatly increases the difficulty of accurate and robust prediction. At the same time, with the theoretical innovation in machine learning, the rise of high-performance computing and big data technologies, deep learning has gradually evolved from an unachievable concept to complex machine learning models that can surpass many models to achieve accurate prediction and large-scale application deployment. Therefore, more and more researchers are concerned on deep learning theory and applications.
The existing shallow learning prediction methods have the characteristics of vulnerability, shallowness, and finiteness. Many shallow learning models are difficult to accurately model sudden traffic events, and incapable of extracting rich features. They often perform well for predictions within a certain period of time, and therefore cannot be utilized to successfully solved the problem of travel time prediction of urban roads. Deep learning methods, however, are good at multi-level feature extraction and can model time-dependent data precisely. However, deep learning also faces the challenges of great training costs and difficulty in hyper-parameter optimization. Therefore, the purpose of this thesis is to establish deep learning models and methods for urban travel time prediction, to compare the representative shallow learning models and methods from the perspectives of stability and accuracy. The main research contents of this thesis are as follows:
1) Conduting in-depth comparative analysis of the development of both shallow and deep learning methods, and establishing a short-term dynamic prediction mechanism based on sliding time windows;
2) Proposing a LSTM-DNN deep learning model for urban road travel time prediction, and implementing classic shallow learning models (e.x., linear regression, ridge regression, LASSO regression, random forest) to perform a comprehensive comparison of prediction performance;
3) In order to construct an optimal deep learning model, this paper innovatively proposes a deep learning hyperparameter optimization method based on grid search, multiple prediction scenarios and multi-group training method. The multi-group training method can avoid randomness problem of the deep learning model. Experiments on 10 different prediction scenarios based on Shenzhen City’s urban road travel time data in 2017 is coducted to obtain comprehensive comparison of model performance for different hyperparameters.
4) A group of ensemble learning models with meta-learners based on shallow learning and deep learning models are established to explore whether the ensemble learning model can integrate the different features of shallow learning and deep learning models to improve the prediction of urban travel time.
Keywords: travel time prediction, deep learning, recurrent neural network, LSTM, hyperparameter optimization, ensemble learning
