پیش بینی زمان رسیدن اتوبوس با استفاده از شبکه های عصبی RBF تعدیل شده با داده های آنلاین Bus Arrival Time Prediction Using RBF Neural Networks Adjusted by Online Data

۱٫ Introduction With the continuous development of the Intelligent Transportation System (ITS), the Advanced Public Transportation System (APTS) and Advanced Traffic Information System (ATIS) become more and more important. Bus arrival time forecasting system belongs to these systems (as shown in Fig.1.). Releasing bus arrival time information to passengers’ mobile devices, helps them to plan their travel time, and to save their waiting time at bus stops. It makes more sense to attract additional passenger flow by providing bus arrival time information to improve the service quality of transit systems. Besides, operators will be able to monitor the execution of schedule, to react instantly and to evaluate the operational efficiency. Bus arrival time predicting method or algorithm designing was considered to be the most complicated part in the former studies. Researches on bus arrival time forecasting started by the end of 1990s which aims to extract bus operating information from vehicle monitoring systems (Lin and Zeng, 1999). With the idea of providing bus arrival information to passengers comes into our sights lately, studies focus on this domain afresh. A case study in Jinan, China was processed by Lin, et al, who proposed two artificial neural network (ANN) models to predict the realtime bus arrivals, based on historical Global Positioning System (GPS) data and automatic fare collection (AFC) system data, which illustrated ANN models are valid to bus arrival time predicting (Lin et al., 2013). Zhou, et al developed an entire system solely relies on the collaborative effort of the participating users and is independent from the bus operating companies instead of referring to GPS enabled location information from specific transit agencies (Zhou et al., 2012). Zhu et al explicitly incorporated the bus stop delays and signalized intersection delays associated with the total travel times of the buses (Zhu et al, 2011). Biagioni, et al developed an online dynamic algorithm of automatic transit tracking, mapping, and arrival time prediction by using smartphones (Biagioni et al, 2011). While Yu, et al used several methods such as support vector machine (SVM), artificial neural network (ANN), k nearest neighbours algorithm (k-NN) and linear regression (LR) as comparisons (Yu et al., 2011). A heuristic method was proposed by Yu, et al which contains two main steps, in which firstly SVM was trained to perceive the historical data and Kalman Filter was applied to import the real-time data (Yu et al., 2008). Nowadays, the stage comes to the implementation and application of such systems from the blueprints with the gigantic development of the electronic, communication, computer software and network engineering since 21th century. Especially profiting from the popularization of smartphones, makes it possible to establish a bus arrival time forecasting system to deliver real-time bus information to both operators and passengers. On the other hand, Automobile Data Recorder and Automatic Vehicle Location (AVL) devices are equipped on bus vehicles as regular equipment, which signify that real-time location of vehicle can be available. Several cities in China have adopted bus arrival predicting systems for bus information offering online (website) or offline (electronic station board), although they can only provide an accuracy by the number of stops instead of terms of time, for example in Dalian, Xiamen and Suzhou, etc. While, can we develop a kind of system with accuracy, stability and simplification that can predict the exact time when will the next bus arrive? In this paper, we propose an approach combining historical data and real-time situation information to forecast the bus arrival time. Firstly, Radial Basis Function Neural Networks (RBFNN) are used to learn and approximate the nonlinear relationship in historical data, so that the results can be given by the trained networks as the basic information references in the first phase. Then, in the second phase, in order to mitigate the influence from discrepancy between historical and real-time data, an online oriented method is introduced to adjust to the actual situation, which means to use the practical information to modify the predicted result of RBFNN in the first phase. Consequently, the result considered to be more dependable can be offered to transit operators, electronic station boards or passengers’ smartphones. Afterwards, the system designing outline is given to summarize the structure and components of the system. We did an experimental study on bus route No.21 in Dalian by deploying this system to demonstrate the validity and effectiveness of this approach. In addition, Multiple Linear Regression model, BP Neural Networks and RBFNN without online adjustment are used in contrast. Results show that the approach with RBFNN and online adjustment has a better predicting performance.