| With the development of economy,traffic congestion is aggravating in the city.The prior development of public bus,as the fundamental solution to traffic congestion,however suffers from disadvantages like low service level and insufficient passenger attraction.Under the background of Intelligent Traffic System(ITS),the intelligentization and informatization of conventional public transportation has greatly improved the reliability of public transportation system and the comfortableness as well as travel efficiency of passengers,it can also reduce congestion and pollution emissions.Travel time is one of the most concerned information related to bus for passengers.Accurate extraction and prediction of travel time is of great significance for both passengers and regulatory authorities.The purpose of this study is to explore the regular pattern of travel time and develop a prediction model so as to help passengers avoid unnecessary waiting and to realize the intelligent control of vehicles and to promote the development of bus first strategy.To start with,the GPS data collected from vehicles is preprocessed,and the bus arrival station is determined according to the longitude and latitude information of the station and vehicle GPS data.Then the bus arrival time and departure time are determined according to the auxiliary information such as speed and regular operation pattern of buses.The travel time is thereby finally determined.After that,the variation characteristics of bus travel time are studied from the dimensions of both whole journey and sections,especially on the similarities and differences between working days and non-working days.The fluctuation coefficient is constructed based on statistical theory and the fluctuation characteristics of bus travel time are explored using the fluctuation coefficient.After extracting the fluctuation coefficient and speed from travel time data,the FCM algorithm is used to cluster the travel time data based on the extracted indexes,the results are compared with common clustering algorithms.Combined with practical experience,the traffic state is divided into saturated flow,free flow and intermediate state.Then,a single station travel time prediction method based on ABC-ELM model is proposed.The proposed ABC-ELM algorithm is compared with the commonly used BP neural network,SVM and ELM models in five scenarios:peak/off-peak,workday/non-workday,sunny day/rainy day,road sections with different use and model running time.Experimental result shows that in peak hour scenario,compared with other algorithms,the ABC-ELM algorithm’s MAPE drops by 7.93%at most,MAE reduces 10.73s at most,and RMSE reduces10.72s at most,its R~2reaches 0.87.The ABC-ELM algorithm also keeps the optimal performance in other three scenarios.In terms of learning time,the average learning time of ABC-ELM algorithm is 110.7s,which is second only to the ELM model(76.5s)and is better than the BP algorithm(128s)and SVM algorithm(255.3s),with its average learning time reduced by 13.5%and 56.6%respectively.In general,the ABC-ELM algorithm proposed in this paper holds strong robustness and better accuracy for single station travel time prediction.Finally,based on the single station travel time prediction model and the conclusion of FCM clustering,a multi-station prediction model based on FCM-ABC-ELM is constructed by introducing the time iteration factor t.The error and applicability of ABC-ELM and FCM-ABC-ELM models for multi-station prediction in short distance and long distance scenarios are compared.The result shows that in short distance scenario,These two method have similar prediction performance in short distance scenarios.In the long distance scenarios,the MAPE,MAE,RMSE and R~2of ABC-ELM algorithm are 12.60%,18.91,22.39,0.88,while9.75%,12.57,15.03,0.92 for FCM-ABC-ELM model.The relative error of the total travel time and cumulative absolute error of ABC-ELM is 9.71%and 283.67s,while 5.30%and 188.66s for FCM-ABC-ELM model.The newly proposed FCM-ABC-ELM algorithm has better prediction accuracy for multi-station travel time prediction in long-distance scenario. |
