Freeway Experienced Travel Time Reliability Prediction Based On Adaptive Congestion Evolution Pattern Recognition

Accurate and reliable travel time prediction provides key data support for road traffic control and travel services.Existing researches are mainly based on two types of ideas to realize travel time prediction,one is instantaneous travel time prediction,and the other is experienced travel time prediction.In comparison,the latter considers the road traffic state evolution dynamic in the modeling process,and usually can obtain more accurate and reliable prediction results.There are two main deficiencies of the existing freeway experienced travel time prediction methods,which limit its key supporting role in decision-making of traffic control and travel services.On the one hand,most current researches are based on aggregated coarse-grained traffic flow data and simple thresholds to realize traffic state identification,resulting in a risk of unreliable identification results,which leads to a large basis between the traffic congestion evolution pattern extracted on this basis and the actual traffic conditions,and further reduce the accuracy of experienced travel time prediction.On the other hand,the current researches focus on the prediction of the mean value of experienced travel time,and ignores the influence of uncertainty factors such as random volatility of traffic flow and data noise on the prediction results during the modeling process,which brings the risk of unreliable experienced travel time prediction,so it cannot meet the requirements of high reliability of travel time information of traffic managers and travelers in real-world conditions.Considering the above problems,relying on unsupervised machine learning and statistical time series analysis technology,combined with high-resolution traffic flow data of freeways,this paper aims to construct a traffic state identification model based on highresolution traffic flow time series clustering,and propose an experienced travel time prediction method based on adaptive congestion evolution pattern recognition,exploring the reliability prediction technology of experienced travel time,so as to provide key decision-making support for freeway traffic control and travel services.The main research work and conclusions of this paper are included as the following three aspects:(1)In terms of traffic state identification,in order to effectively capture the short-term fluctuation characteristics and trends information in fine-grained traffic flow series(time intervals of 30s,1min,etc.),established the freeway traffic state identification model.The Dynamic Time Warping(DTW)algorithm is used to measure the similarity between finegrained traffic flow time series,and determine the number of optimal traffic state clusters based on the traffic state of key sections of the freeway corridor.Based on the fuzzy C-means clustering(FCM)to realize reliable traffic state identification of off-line sections;in order to overcome the frequent switching of traffic states during a short time period in the online identification results,the online traffic state identification results are effectively smoothed by a method based on the dynamic programming algorithm.The rationality of the constructed model is evaluated and verified by using high-resolution traffic flow data with a time interval of 1 minute.The results show that the constructed model can obtain reliable and reasonable traffic state identification results,thus laying a solid foundation for accurate and reliable traffic congestion evolution pattern recognition in freeway corridors.(2)In terms of experienced travel time prediction,a freeway experienced travel time prediction method based on adaptive congestion evolution pattern recognition is proposed.The variational auto-encoder(VAE)algorithm is used to reduce the dimensionality of the traffic speed spatiotemporal map,to effectively reduce the modeling complexity,the k-means algorithm is used to cluster the speed spatiotemporal map after dimensionality reduction,and strategies of traffic congestion pattern recognition and online search of the spatiotemporal evolution pattern are design,to realize the online prediction of experienced travel time.The proposed method is compared with other two types of benchmarking methods(instantaneous travel time prediction method and historical average travel time prediction method),and the results show that:during the non-congested situation,the prediction effectiveness of the proposed method and the instantaneous travel time prediction method is close,and they are better than the historical average travel time prediction method;during the congestion situation,the proposed method is significantly better than the other two types of benchmark methods,and is closer to the real experienced travel time.The predicted average MAE and MAPE are 2.5minutes and 6.7%,respectively.The excellent potential of the proposed method for decision-making in real freeway applications is verified.(3)In terms of experienced travel time reliability prediction,a method of freeway experienced travel time reliability prediction based on the Family Generalized Autoregressive Conditional Heteroskedasticity(FGARCH)model is proposed.The experienced travel time residual time series are input,and the mean equation is constructed based on the Autoregressive Integrated Moving Average(ARIMA)algorithm to predict the mean value of the residual,then the variance equation is constructed based on the FGARCH algorithm to predict the volatility(confidence interval)of the residual.Finally,the confidence interval of the predicted experienced travel time is obtained by adding the predicted mean value and confidence interval of the predicted residual after recovery.The proposed method was evaluated and validated with four evaluation indexes,including kick-off percentage(KP),average confidence interval length(ACL),volatility mean absolute error(VMAE)and directional accuracy(DA).The results show that t The proposed method can predict reasonable volatility intervals of the experienced travel time,thereby effectively improving the reliability of traffic control and decision-making of travel services;the average KP value of the model is below 8%during the congested and non-congested situations,and the average ACL value is less than 11.5 minutes during the congestion situation.