基于多任务学习的高速路网交通流动态演变与预测研究

项目名称： 基于多任务学习的高速路网交通流动态演变与预测研究

项目编号： No.61473006

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 谢昆青

作者单位： 北京大学

项目金额： 83万元

中文摘要： 针对目前我国高速公路快速成网的现状，从网络角度分析省级高速路网交通流的时空演变规律。基于车辆进出(收费)数据和路段线圈数据，采用多任务学习范式，深入分析和在线挖掘高速路网实时交通流数据，构建适用于高速路网交通流分析的多任务特征学习模型，扩展多任务学习方法的时空动态性；研究常态及非常态交通流关联模式和动态演变过程，利用关联关系变化构建交通事件识别模型，探究高速路网交通关联演变机理；研究多任务并行学习方法和任务分组策略，实现高效准确的路网交通流实时预测；研究多任务关联补偿机制，提高预测模型在数据延迟或缺失情况下的鲁棒性和实用性；研发面向高速公路管理需求的原型系统，可视化展示交通流演变过程。课题将从理论层次上系统理解网络条件下高速公路各组成部分之间的联系和影响，通过对交通流演变规律的系统阐释和交通流变化的准确预测，为科学有效的高速公路运营监控、综合管理、应急处置和信息服务提供基础理论和模型支持。

中文关键词： 高速路网；演变预测；关联分析；多任务学习

英文摘要： With the rapid network formulation of highway system in China, this proposal focuses on analyzing the spatial-temporal patterns of traffic flow in highway from the perspective of network. This study will base on toll data of entrance-exit stations and inductive loop data on the road, apply the paradigm of of multi-task learning in analyzing and online mining real-time traffic flow data of highway system; establish multi-task feature learning model for traffic flow analyzing in highway system, extend the spatial-temporal dynamics of approaches of multi-task learning; investigate the relation patterns and evolution process of highway traffic flow under normal and abnormal traffic conditions, establish abnormal traffic flow recognitio model based on the variation of relation patterns, describe the mechanism of relation and evolution of highway traffic flow; investigate the parallel speeding method and the grouping approach of multi-task learning, implement efficient and effective real-time traffic flow prediction in highway system; investigate the mechanism of compensation based on relation, improve the robustness and reliablity of predition model under data delay or data missing; develop prototype system for the demand of highway management, visualize the process of traffic flow evolution. This proposal will help us understand the relation between each part in the highway system from the theoretical level. From the systematic explanation of the evolution pattern of traffic flow and the accurate predition of traffic flow, this study will lay a solid theoretical foundation for effective monitoring, management and emergence response of highway system.

英文关键词： highway network;evolution and prediction;correlation analysis;multitask learning