In order to estimate traffic flow a Bayesian network BN model using prior link flows is proposed.This model sets link flows as parents of the origin-destination OD flows. Under normal distribution assumptions the mode...In order to estimate traffic flow a Bayesian network BN model using prior link flows is proposed.This model sets link flows as parents of the origin-destination OD flows. Under normal distribution assumptions the model considers the level of total traffic flow the variability of link flows and the violation of the conservation law.Using prior link flows the prior distribution of all the variables is determined. By updating some observed link flows the posterior distribution is given.The variances of the posterior distribution normally decrease with the progressive update of the link flows. Based on the posterior distribution point estimations and the corresponding probability intervals are provided. To remove inconsistencies in OD matrices estimation and traffic assignment a combined BN and stochastic user equilibrium model is proposed in which the equilibrium solution is obtained through iterations.Results of the numerical example demonstrate the efficiency of the proposed BN model and the combined method.展开更多
物理信息深度学习(physics-informed deep learning, PIDL)是一种将深度学习与物理学先验知识相结合的新兴范式,该范式在智能交通领域,尤其在交通状态估计应用中,展现出了巨大潜力。为进一步优化物理信息深度学习模型在交通状态估计问...物理信息深度学习(physics-informed deep learning, PIDL)是一种将深度学习与物理学先验知识相结合的新兴范式,该范式在智能交通领域,尤其在交通状态估计应用中,展现出了巨大潜力。为进一步优化物理信息深度学习模型在交通状态估计问题上的准确度与收敛速度,构建了一个结合Aw-Rascle宏观交通流模型的物理信息自适应深度学习模型(physics-informed adaptive deep learning with Aw-Rascle, PIAdapDL-AR),依据有限与局部的交通检测数据,实时准确估计全局交通流状态。主要的改进包括两部分,一是在PIDL框架中的物理部分引入高阶Aw-Rascle交通流模型作为物理约束条件,引导并规范神经网络的训练过程;二是在神经网络部分融合自适应激活函数,替代固定的非线性激活函数,以动态优化神经网络性能。基于NGSIM数据集生成模拟的固定检测器数据和移动检测器数据,进行实验以验证模型有效性。实验结果表明:在不同覆盖率的固定检测数据场景下,PIAdapDL-AR的相对误差相比于基线模型PIDL-LWR降低了34.38%~45.24%;在不同渗透率的移动检测数据场景下,PIAdapDL-AR的相对误差相比于PIDL-LWR降低了18.33%~34.95%;融合自适应激活函数的PIAdapDL-AR的收敛速度优于配置固定激活函数的PIDL-AR,且收敛速度和估计精度均随着自适应激活函数中比例因子的增大而提升。展开更多
基金The National Natural Science Foundation of China(No.51078085,51178110)
文摘In order to estimate traffic flow a Bayesian network BN model using prior link flows is proposed.This model sets link flows as parents of the origin-destination OD flows. Under normal distribution assumptions the model considers the level of total traffic flow the variability of link flows and the violation of the conservation law.Using prior link flows the prior distribution of all the variables is determined. By updating some observed link flows the posterior distribution is given.The variances of the posterior distribution normally decrease with the progressive update of the link flows. Based on the posterior distribution point estimations and the corresponding probability intervals are provided. To remove inconsistencies in OD matrices estimation and traffic assignment a combined BN and stochastic user equilibrium model is proposed in which the equilibrium solution is obtained through iterations.Results of the numerical example demonstrate the efficiency of the proposed BN model and the combined method.
文摘物理信息深度学习(physics-informed deep learning, PIDL)是一种将深度学习与物理学先验知识相结合的新兴范式,该范式在智能交通领域,尤其在交通状态估计应用中,展现出了巨大潜力。为进一步优化物理信息深度学习模型在交通状态估计问题上的准确度与收敛速度,构建了一个结合Aw-Rascle宏观交通流模型的物理信息自适应深度学习模型(physics-informed adaptive deep learning with Aw-Rascle, PIAdapDL-AR),依据有限与局部的交通检测数据,实时准确估计全局交通流状态。主要的改进包括两部分,一是在PIDL框架中的物理部分引入高阶Aw-Rascle交通流模型作为物理约束条件,引导并规范神经网络的训练过程;二是在神经网络部分融合自适应激活函数,替代固定的非线性激活函数,以动态优化神经网络性能。基于NGSIM数据集生成模拟的固定检测器数据和移动检测器数据,进行实验以验证模型有效性。实验结果表明:在不同覆盖率的固定检测数据场景下,PIAdapDL-AR的相对误差相比于基线模型PIDL-LWR降低了34.38%~45.24%;在不同渗透率的移动检测数据场景下,PIAdapDL-AR的相对误差相比于PIDL-LWR降低了18.33%~34.95%;融合自适应激活函数的PIAdapDL-AR的收敛速度优于配置固定激活函数的PIDL-AR,且收敛速度和估计精度均随着自适应激活函数中比例因子的增大而提升。
