Traffic Speed Estimation And Prediction Under Normal And Abnormal Conditions

Traffic originates from the necessity of travel and mobility in people's daily life,so traffic has become an essential part of connecting residents' activities.Improving the traffic environment and realizing smooth and safe traffic have become a primary national concern for people's livelihood issues.At present,major cities are vigorously developing intelligent transportation systems,and accurate traffic state estimation and prediction form the foundation of smart traffic management and control.Although a large number of traffic flow detection devices are deployed in the city,it is still unable to achieve the full-time and spatial collection of traffic data,and various random factors often interfere with traffic flow.Data loss,data fusion,and data noise have become challenges to achieve accurate traffic flow prediction.At the same time,the parameters of common traffic flow estimation and prediction models do not change once calibrated.Most of the existing research focuses on traffic flow prediction under normal conditions,but the research on abnormal events is insufficient.As a vital skeleton of the urban road network,the urban expressway is often affected by the frequent merging,interweaving,and diverging of vehicles at the on/off-ramps.Compared with the freeway system,the evolution of traffic states on urban expressways are more complicated.This thesis proposes an adaptive rolling smoothing method,which combines microwave and license plate recognition data to estimate and predict the traffic flow of an individual segment of an urban expressway.The spatio-temporal speed field is reconstructed by nonlinear interpolation,and the experienced travel time of an individual vehicle is estimated based on the virtual trajectory algorithm.The parameters of default filter,global optimization(static parameter setting),and rolling horizon(dynamic parameters)are compared.The adaptive rolling smoothing method can dynamically adjust the model parameters at the end of the update time horizon,solve the problem that the fixed filter parameters cannot dynamically represent the characteristics of traffic flow in the existing methods,and better capture the dynamic evolution process of traffic congestion generation,propagation and dissipation.Considering that traffic detection data is a complex non-linear time series including noise,if it can be decomposed into more periodic components,then the prediction accuracy of the model can be improved.This thesis proposes a DeepEnsemble model framework that integrates ensemble empirical mode decomposition(EEMD)with a three-dimensional convolution neural network(3D CNN)to realize decomposition,prediction,reconstruction,and fusion of urban expressway network-level traffic flow.The original time series is decomposed into several intrinsic mode functions and one residual by using EEMD.A three-dimensional input tensor is constructed by placing the historical steps in the time dimension,the detectors in the spatial dimension,and the decomposed sub time series in the depth dimension.3D CNN is used to extract traffic flow features from the three dimensions of time,space,and depth.The model considers the influence of external features and historical information,fully extracts the temporal and spatial correlations of traffic flow,and realizes multiple detectors and multi-step prediction of a large-scale urban expressway network.Traffic flow is often affected by unexpected incidents.However,traffic flow prediction models trained based on normal conditions may not generate good results under abnormal conditions,so it is necessary to develop a more robust algorithm to improve the accuracy of traffic flow prediction under abnormal conditions.Inspired by the idea of ensemble learning,a better result can be generated by integrating multiple sub-optimal prediction results.In this paper,a general multi-model ensemble learning framework is proposed,and a two-level uncertainty ensemble model based on gradient boosting regression tree and Lasso is constructed.The model focuses on the uncertainties of model structure and model parameters,and improves the accuracy of traffic flow prediction under abnormal conditions.By comparing the prediction results under normal and abnormal traffic conditions,the effectiveness of the proposed two-level ensemble model is verifiedThe research scope of this paper is expanded from traffic flow estimation and prediction of an individual road segment of urban expressway with heterogeneous data to traffic flow prediction for the large-scale urban expressway network.The research issues are expanded from traffic flow prediction under normal traffic conditions to abnormal traffic conditions.Through the construction of a variety of traffic flow prediction models,this thesis has achieved an accurate prediction of traffic flow in different areas and under different conditions.