Travel time reliability assessment has been widely used in recent years to evaluate the performance of transportation networks and measure the operation level of transportation systems. Weather, as one of the important factors influencing travel time reliability, affects the relationship between the supply and requirement of urban road networks. Considering the traffic characteristics under different traffic conditions, a study on the influence of weather on travel time reliability under different conditions is proposed to predict the probability of travelers completing their trips within the expected time under different weather conditions. Based on the urban road network data and cab trajectory data of Harbin city, this paper correlates the floating vehicle location with the road network information through a hidden Markov model to reduce the influence of vehicle trajectory errors on the calculation results of path travel time. To analyze the entire distribution of extreme travel time and its impact on the reliability of travel time under various traffic situations, it captures the tail features of the travel time distribution based on extreme value theory. Then, to increase the predictability of each quantile, it combines a deep-learning LSTM model and a quantile regression model to create a probabilistic travel time prediction model utilizing combined layers. The proposed model is compared with the linear quantile regression and neural network quantile regression models, and the model is evaluated in terms of point prediction results and probabilistic prediction results, respectively, to ensure the accuracy of predictions from the model. As a result, the prediction accuracy of the model in this paper is greatly improved, and the degree of violating quantile constraints is greatly reduced.

Travel time prediction is one of the most important parameters to forecast network-wide traffic conditions. Travelers can access traffic roadway networks and arrive in their destinations at the lowest costs guided by accurate travel time estimation on alternative routes. In this study, we propose a long short-term memory (LSTM)-based deep learning model, deep learning on spatiotemporal features with Convolution Neural Network (DLSF-CNN), to extract the spatial–temporal correlation of travel time on different routes to accurately predict route travel time. Specifically, this model utilizes network-wide travel time, considering its topological structure as inputs, and combines convolutional neural network and LSTM techniques to accurately predict travel time. In addition to their spatial dependence, both coarse-grained and fine-grained temporal dependences are fully considered among the road segments along a route as well. The shift problem is formulated in the coarse-grained granularity to predict the route travel time in the next time interval. The experimental tests were conducted using real route travel time obtained by taxi trajectories in Harbin. The test results show that the travel time prediction accuracy of DLSF-CNN is above 90%. Meanwhile, the proposed model outperformed the other machine learning models based on multiple evaluation criteria. The RMSE (Root Mean Squard Error) and R² (R Squared) increased by 18.6% and 22.46%, respectively. The results indicate the proposed model performs reasonably well under prevailing traffic conditions.

It has been recognized that significant travel time estimation errors may be introduced using low-resolution GPS-based floating car trajectory data traditionally. Very few studies have been conducted to concentrate on spatial–temporal relationship identification among travel time measurements. In this study, an attention-based spatial–temporal graph convolutional networks on low-resolution data (AGCN-LR) approach was proposed to estimate more accurate travel time in urban traffic roadway networks using low-resolution GPS-based data measured by floating cars. Specifically, three models were developed in this AGCN-LR approach. Hour, day, and week were used to model the dynamic relationship among spatial–temporal traffic flow attributes, respectively. The same structures were adopted for these three models. Two spatial-temporal block (ST-block) models and one temporal convolutional model were included. Furthermore, one spatial graph convolutional model and one temporal attention mechanism model were embedded in a ST-block. AGCN-LR not only improved the efficiency and accuracy of travel time estimation through the framework optimization training process in a spectrum convolution network but also combined the three temporal components. The final estimation value was formed afterward. Experimental tests were conducted using the real data set from low-resolution floating car data in Harbin, China, in 2017. Results indicated that AGCN-LR outperforms the other state-of-the-art algorithms by reducing estimation mean absolute error (MAE) by about 50 s when it captured the relationship among dynamic spatial and temporal data from the data set. The AGCN-LR approach demonstrated great potential to become one of the important urban network-wide traffic management tools using low-resolution floating car data.