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,PIAda-pDL-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.
Physics-Informed Adaptive Deep Learning-based Traffic State Estimation
Physics-informed deep learning(PIDL)is an emerging paradigm that combines deep learning with the physics prior knowledge.This paradigm has shown great potential in the field of intelligent transportation,especially in the application of traffic state estimation.To further optimize the accuracy and convergence speed of PIDL models in traffic state estimation problems,this paper proposes a physics-informed adaptive deep learning with Aw-Rascle(PIAdapDL-AR)model that combines Aw-Rascle to accurately estimate the global traffic flow state based on limited and local traffic detection data.The main improvements include two parts.Firstly,in the physical part,a high-order Aw-Rascle traffic flow model is introduced as a physical constraint to guide and standardize the training process of the neural network;The second is to integrate adaptive activation functions in the neural network,replacing fixed nonlinear activation functions,to dynamically optimize the performance of the neural network.The NGSIM dataset isusedto generate simulated fixed detector data and mobile detector data,and experiments are conducted to verify the effectiveness of the model.The experimental results show that in fixed detection data scenarios with different loop number,the relative error of PIAdapDL-AR is reduced by 34.38％to 45.24％compared to the baseline model PIDL-LWR;Besides,in different penetration rates of mobile detection data scenarios,the relative error of PIAdapDL-AR is reduced by 18.33％to 34.95％compared to PIDL-LWR;In addition,the convergence speed of PIAdapDL-AR,which integrates adaptive activation functions,is better than that of PIDL-AR with fixed activation functions,and both the convergence speed and estimation accuracy can be improved with the increase of the scaling factor in the adaptive ac-tivation function.
intelligent transportationtraffic state estimationphysics-informed deep learningtraffic flowneural networks
万方数据
维普
