Expressway Traffic State Estimation Based On Physics-informed Machine Learning And Congestion Evolution Analysis

The evolution analysis of expressway traffic congestion provides key decision-making support for scientific and reliable expressway management and control.Scholars at home and abroad have carried out in-depth exploration on the evolution analysis and modeling of highway traffic congestion,and achieved fruitful research results.In spite of this,there are still two major deficiencies in the current research,which make it difficult to play a key supporting role in road traffic control.Firstly,in order to obtain refined traffic operating state parameters,researchers proposed a traffic state estimation method based on model driven and data driven.The model driven method builds an estimation model based on the analytical relationship of traffic state variables,which has the advantages of high data efficiency and strong interpretability in the case of small samples.However,due to the relatively large number of simplifying assumptions and being too strict,the model estimation performance usually deviates greatly from the actual traffic state.The data-driven method is mainly based on artificial intelligence algorithms such as neural networks to build estimation models and learn traffic state evolution rules from a large number of real data.It has the advantage of high generalization accuracy,but poor data efficiency and interpretability.It is still an open topic to explore the traffic state estimation method with high generalization accuracy and strong interpretation.Secondly In terms of traffic congestion evolution analysis,existing research mainly relies on artificially setting thresholds to judge traffic congestion status,lacks an automatic mechanism,and the congestion evolution analysis index is relatively single,which cannot satisfy a comprehensive and systematic expressway congestion evolution analysis.Considering the above problems,this paper uses the theoretical knowledge and technology of traffic flow model,neural network and unsupervised clustering to study and explore the highway corridor traffic state estimation and congestion evolution analysis.The research results and conclusions obtained in this paper are summarized as follows:(1)Traffic state estimation:a physical information-based machine learning traffic flow state estimation method is proposed,which is based on the integration of two approaches,model-driven and data-driven.The method features the introduction of time,space and velocity physical information into the framework of processing time-series LSTM deep learning models,learning the relationship between physical variables and extracting the bias values of velocity against time and space through the neural network back-propagation calculation process,substituting the bias values into a first-order non-linear traffic flow conservation law model as a physical constraint for training the LSTM model,guiding and standardising the neural network learning process.Given 7 different numbers of detection sites to verify the estimation performance of the model for different sample sizes,three metrics,MAE,RMSE and MAPE,are chosen for evaluation and 2 benchmark models,linear traffic flow based machine learning and pure LSTM,are introduced for comparison.The experimental results illustrate that the constructed model has the best estimation results compared to the benchmark model,especially the comparison estimation results are more significant for the case of only 6 detector samples(a small number of samples),and it improves the estimation accuracy by 31.17%,11.65%and 18.71%respectively compared to the LSTM.(2)For recurrent traffic congestion bottleneck identification.Firstly,the K-means algorithm is used to cluster the traffic history data,and the segmentation speed is set according to the clustering result,so that the traffic congestion status can be discerned;secondly,the congestion probability of each spatio-temporal data point is calculated,and the traffic congestion type is studied and judged based on the random congestion map;then,the GTBC algorithm is used to search the connectivity of the data points in the recurrent spatio-temporal congestion area,so as to judge the number of congestion areas.Based on this,the automatic identification of recurrent traffic congestion bottlenecks is carried out for both weekday and weekend modes according to the characteristics of traffic congestion bottlenecks.The experimental results show that:① on weekdays,there are four or more congestion bottlenecks in the corridor,with longer duration;② on weekends,there is only one more obvious congestion bottleneck,and even if the others are identified,they have a small congestion range and short duration.(3)The spatio-temporal evolution of recurrent traffic congestion:Firstly,the spatial and temporal congestion area contour points are extracted to characterise the quantification of the spatial and temporal evolution of congestion;secondly,indicators such as the start time of traffic congestion,the end time of traffic congestion,the spatial location of the start of congestion and the spatial location of the end of congestion are calculated based on the information of congestion data points to effectively portray the spatial and temporal evolution of traffic congestion.Then,the distribution characteristics of the data points corresponding to the maximum and minimum speed of the congestion area contour points are fitted to achieve the propagation speed of traffic congestion in the diffusion and dissipation phases.Finally,the spatio-temporal map of traffic flow is combined to obtain the traffic flow corresponding to each data point,so that the delay of each data point can be calculated,thus realising the total delay of each spatio-temporal congestion area.