基于大数据深度学习的背压分布式过饱和信号控制方法

Worsening traffic congestion in urban areas result in a growing number of signalized intersections operates in oversaturated conditions, and the duration for the congestion condition also has an expanding trend. It has become a very serious social problem in most big cities all over the world now. With the increasingly rich of multi-view big data and the increasingly mature of data mining technologies, we want to present a new traffic control method based on the back-pressure technology using big data deep learning in the new project. It will break the traditional signal control with a centralized and passive mode, and owns the following two major advantages comparing to the existing ones: (1) it is an distributed mode and can be used in a large area; and (2) the traffic flow data needed in the new method are provided by deeply learning, which can ensure the precision of the basic parameters. Firstly, we complete the data quality inspection and repair the abnormal data. Considering the characteristics of non-uniformity and sparsity for the multi-view traffic big data, a sparse optimization algorithm should be advanced using the evolutionism idea. Meanwhile, a completion method for missing data and a fusion algorithm for different kinds of data are also needed by mining of the association rules among different sources of traffic data together with analyzing the relationships between different traffic flow parameters (such as traffic volume, speed, density). After separating the individual panoramic or partial trajectory information into different trips, the data about travel route and node OD can be obtained. Then, with a long time range of these two kinds of data mentioned above, a real-time prediction method of the individual route choice behavior based on the deep learning will be established, which is helpful to study the mutual feed model between the state evolution and the route choice. Thirdly, we will tease the control strategy of distributed mode with pressure-back and define the starting condition based on the macroscopic basic diagram. With the actual requirements, the back pressure coefficient estimation method need to be reconstructed under the dynamic route adjustment, and then the optimization methods for the boundary and internal nodes will be advanced respectively. At last, the advancement of the new method will be verified based on simulations and actual measurements. The results of the project can expand the signal control strategy system and provide a reference method to alleviate the urban traffic congestion.

针对常态化与区域化的拥堵态势，利用日益丰富的多视图大数据和日益成熟的数据挖掘技术，打破传统集中、被动式信号控制模式的束缚，借鉴路由调度中背压控制思想，建立一套基于大数据深度学习的背压分布式过饱和信号控制方法。首先完成数据质量检验和预处理，针对大数据视图非一致性和高维稀疏性，构造基于进化思想的多目标稀疏优化方法，利用关联规则挖掘及交通流特性参数模型解析，完成数据补全与融合；其次，分离出行个体全景或部分轨迹信息，获取出行路径及节点OD，并扩展时间范围推导基于深度学习的出行个体路径选择行为实时预测模型，构建状态演变与路径选择之间的互馈机制；再次，梳理背压分布式信号控制策略，基于宏观基本图界定启动条件，重构动态路径调整下的背压系数测算方法，建立拥堵区域边界节点信号调整算法和内部节点分布式控制算法，并基于仿真和实测完成方法验证。项目成果应能拓展信号控制策略体系，为缓解城市交通拥堵提供一种参考思路。

面向日益严峻的城市交通拥堵问题，项目以丰富的交通大数据为基础，研究一种洞察交通时空分布及拥堵本源的数据挖掘方法，进而构建了一套可以主动预防排队溢流发生的信号控制方法，主要研究成果总结如下：.1.构建了卡口、微波等多源交通大数据的质量控制与预处理方法，建立了个体出行链提取及补全机制，并融合出行特征的时空分布特点实现了出行链的自动打断与节点OD自动生成；研究节点OD的短时分布特征表征方式，辨识拥堵路径与拥堵车流；搭建节点OD的长时规律挖掘方法，建立出行目的的预测模型，界定出行需求结构分布，为针对性信号调节奠定数据基础；.2.信号控制是针对未来一段时间交通流的路网时空资源调控，需要精准把握交通态势的演化规律。交通系统具有很强的时空相关性，并与外界的天气、日期、事故等要素密切相关；本项目首先基于深度学习技术搭建了交通流模式与外部要素的耦合关系，提出了数据集精化方法；探究了交通流长短时模式的特征提取及融合方法，建立了面向多尺度、多场景的交通流态势时空推演模型。.3.基于多源交通大数据建立了区域宏观基本图的融合生成方法，借助三相交通流策略界定了拥堵控制的状态阈值；搭建了基于空间预测控制的路段交通状态互馈演化模型，探究了基于车道组的路网背压系数测算方法，并以路径间和路径内的背压系数双重均衡为目标，面向固定式相位相序结构和可变式相位相序结构，完成了适应不同场景的排队溢流主动预防信号控制方法。.项目成果为预防拥堵蔓延提供了一种新的思路，为未来的信号控制系统升级提供了一种参考方法，对于缓解城市交通拥堵具有一定意义。结合项目研究已发表SCI检索论文21篇，申报国家发明专利6项，参与行业标准2项，获中国智能交通协会科技进步一等奖1项、中国电子学会科技进步二等奖1项；相关成果已联合企业在深圳市进行了示范应用，取得了良好控制效果。