基于多源异构大数据融合的区域交通拥堵演化规律及其预报方法研究

With the double-rapid expansion of total travel volume and travel scope of the car in China, new problem of highway traffic congestion has appeared, that is, traffic congestion has spread from a localized bottleneck to a regional network. Therefore, under the background of big data and artificial intelligence, this project comprehensively employs multiple disciplines of knowledge, such as traffic science, management science, system science and information science, to deeply understand the Spatial-temporal distribution and evolution of regional road network traffic, and to implement traffic forecasting. The studies include: (1) Aim at solving the heterogeneous and sparse problem of multi-source traffic big data, the common structures of multi-source traffic data are extracted and the dynamic tensor stream fusion model is newly designed. And then, regional road network microscopic traffic flow parameters are reconstructed, and its basic law of evolution is revealed. (2) Aim at regional road network characteristics, macroscopic fundamental diagram (MFD) theory is introduced to analyze the influence mechanism of many factors such as road network property and traffic behavior heterogeneity on traffic status. The methods for dynamic traffic state identification and sub-control region division are presented as well. After that, the appearance, evolution and disappearance laws of regional traffic congestion are explored. (3) Aim at the forecast accuracy, a high-dimensional eigenvector of integrating the microscopic traffic flow parameters and macro state parameters is constructed. And then, a bi-level optimization forecast model based on the long-short term memory network and support vector regression is built to enhance the forecasting stability of regional traffic congestion. The implementation of prototype system development and traffic practice of this project, can expend the network traffic flow theory, and develop highway intelligent transportation system, and finally provide theory and method supports for promoting the level of relieving traffic congestion and traffic management.

随着我国汽车出行总量和范围快速双扩张，高速公路交通拥堵呈现由局部瓶颈向区域路网蔓延新问题。为此，本项目综合交通科学、管理科学、系统科学、信息科学等多学科知识，在大数据和人工智能背景下，深入理解区域路网交通时空分布及演化规律并实施交通预报：（1）提取数据共性结构，创新设计动态张量流模型，解决异构、稀疏多源大数据融合问题，重建区域路网微观交通流参数并揭示其演变规律；（2）研究路网属性、交通行为异质性等多因素对交通状态的影响机制，基于宏观基本图理论提出交通状态动态辨识及控制子区域划分方法，探索区域路网交通拥堵产生、演化与消散机理；（3）构造集微观交通参数与宏观状态参数于一体的高维特征向量，基于长短时记忆网络和支持向量回归设计双层优化预报模型，提升区域路网交通拥堵状态预报精准性。通过系统研发和交通实践，为丰富网络交通流理论体系、构建智能交通系统、提高缓解拥堵和交通管理水平，提供基础理论与方法支持。

随着我国汽车出行总量和范围快速双扩张，高速公路交通拥堵呈现由局部瓶颈向区域路网蔓延新问题。为此，本项目深入探究区域路网交通流时空分布及演化规律并实施交通预报：.（1）基于多源交通数据重构区域路网交通流参数及微观模型方面：基于NGSIM数据研究车辆行驶轨迹特性，提出了一种融合CNN、GRU、注意力机制的模型，精准识别车辆换道意图。基于长距离雷达检测的车辆轨迹数据，提出了一种基于LSTM的跟驰速度预测模型，发现变速意愿随速度升高而降低。基于路网覆盖强度和覆盖率分析浮动车数据的最小样本需求量，提出了基于截断核范数的低秩张量分解模型，解决了稀疏数据下交通流分布时空演化规律的发掘难题。.（2）动态辨识交通状态及划分控制子区域，探索区域路网交通拥堵产生、演化与消散机理方面：面向复杂城市道路网络，提出了一种结合蒙特卡罗的间接度量方法，精准识别关键路段。针对城市区域路网OD缺少问题，基于MFD提出了一种FR-PID区域边界控制方法，减少了区域累计延误。针对高速路与快速路毗邻区域路网结构异质问题，提出了一种基于密度峰值优化的NCut算法，将异质性较大路网划分为多个同质性较强的子区域，实现了区域路网交通状态精准识别。.（3）构造宏微观参数于一体的高维特征向量，研发交通预报原型系统，提升预报精准性方面：针对高速公路与快速路毗邻区域的非常发性拥堵，提出了一种GRU优化算法，精准预测车辆行驶速度。针对高速服务区交通流波动大的特征，提出了一种基于STL-OMS的组合交通流预测模型。针对高速收费站的常发性拥堵，构建了BOA-LSTM组合预测模型，并提出了一种车道开闭配置方法，提高了预测精度且减少了收费站运营成本。融合高速公路ETC门架和RTMS等多源数据预测区间行程时间，预测精度明显提高。.本项目基于多源交通异构大数据分析区域路网交通流特性，重建微观交通流参数及模型，识别宏观路网交通状态，构建交通拥堵组合预报模型，实现精准预测，为构建智能交通系统、提高缓解拥堵和交通管理水平提供理论与方法支持。