Comparison of Model Based and Machine Learning Approaches for Bus Arrival Time Prediction

3 The provision of accurate bus arrival information is critical to encourage more people to use 4 public transport and alleviate traffic congestion. Developing a prediction scheme for bus travel 5 times can provide such information. Prediction schemes can be data driven or may use a 6 mathematical model that is usually less data intensive. This paper compares the performance of 7 two methods – one being the data driven Artificial Neural Network (ANN) method and the other 8 being the model based Kalman filter method, with regards to predicting bus travel time. The 9 performances of both methods are evaluated using data collected from the field. It was found that 10 the ANN based method performed slightly better in the presence of a large database but the 11 Kalman filter method will be more advantageous when such a database is not available. 12 TRB 2014 Annual Meeting Paper revised from original submittal. Vivek, Anil, Lelitha and Shankar 2 Comparison of Model Based and Machine Learning Approaches for 1 Bus Arrival Time Prediction 2 INTRODUCTION 3 In recent times, traffic congestion has been increasing in Indian cities due to rapid changes in 4 urbanization, economy levels, vehicle ownership and population growth that has lead to several 5 negative impacts such as delays, pollution, etc. To overcome these problems, there is a need to 6 provide more facilities by infrastructure expansion such as Bus Rapid Transit Systems (BRTS) 7 (Ahmedabad), and Metro Rail Systems (Delhi, Hyderabad, Chennai, Bangalore). However, 8 infrastructure expansion alone cannot meet the vehicular growth, and hence, there is a need to 9 explore better traffic operations and management systems. One of the major challenges of traffic 10 management in most of the developing countries such as India is due to heterogeneous traffic, 11 which comprises both motorized and non-motorized vehicles with diverse vehicular 12 characteristics. The motorized or fast moving vehicles include passenger cars, buses, trucks, auto 13 rickshaws and motor cycles, whereas the non-motorized or slow moving vehicles include 14 bicycles, cycle rickshaws and animal drawn carts. The use of Intelligent Transportation Systems 15 (ITS) for operation and management of traffic is a better option that is gaining interest in recent 16 years. Attracting more travelers towards public transportation system is one way to reduce 17 congestion, which comes under Advanced Public Transportation Systems (APTS). APTS is one 18 of the functional areas of ITS that can help to attract more people towards public transport. 19 Predicting accurate bus travel times and providing reliable information to passengers is a popular 20 APTS application. However, the information provided to passengers should be reliable; 21 otherwise customers may reject the system due to lack of reliability(1). The reliability of such 22 information being provided depends on the prediction technique and the input data used for the 23 same. 24 There are many studies on prediction of travel time using various techniques such as 25 historical and real-time approaches, statistical techniques, machine learning techniques and 26 model-based techniques. However, there are only limited studies under heterogeneous traffic 27 conditions such as the one existing in India. Machine learning techniques such as Artificial 28 Neural Networks (ANN) and Support Vector Machines (SVM) are commonly used to predict 29 travel time because of their ability to solve complex non-linear relationships. ANN has proved to 30 be one of the most effective tools for pattern recognition across different sets of problems. 31 Considering anomalies in datasets, which are true for travel time across a particular stretch of 32 road, ANN seems to be a suitable candidate for prediction. This study attempts the short term 33 Bus Travel Time Prediction (BTTP) by developing a neural network model taking most 34 correlated previous trips of same day and same week in to account. In this study, a particular 35 section of the road is divided into different subsections and the network is trained separately for 36 each subsection. The disadvantage is that these types of techniques need a large amount of data 37 to train the system. 38 Model-based techniques, on the other hand, will use models that capture the dynamics of 39 the system by establishing mathematical relationship between variables. In this study, equations 40 that can characterize the evolution of travel time over space are used. Many model-based studies 41 use the Kalman Filtering Technique (KFT) for estimation. Advantages of this approach are it's 42 limited data requirement and suitability for real time implementation. In this study, a model43 based approach that uses data from just two previous buses will be compared with the ANN 44 technique. 45 TRB 2014 Annual Meeting Paper revised from original submittal. Vivek, Anil, Lelitha and Shankar 3 Thus, the objective of the present study is to predict the travel time of buses under Indian 1 traffic conditions by using ANN and a model-based approach using KFT by providing the 2 appropriate inputs to the models. The study compares the performance of a data driven approach 3 with a model based method that requires only minimal data for bus travel time/arrival time 4 prediction. 5 LITERATURE REVIEW 6 Since the present study focuses on the use of ANN and model based approaches for the bus 7 arrival prediction, some of the previous works in this area are reviewed. Jeong et al. (2) reported 8 a bus travel time prediction model based on ANN taking into account arrival time, dwell time 9 and schedule adherence. Wang et al.(3) used support vector regression taking departure time 10 from the stops as inputs to reflect the traffic conditions. The developed model also used historical 11 bus travel time data, parameters of traffic conditions along bus route, and route specific 12 parameters to predict future bus travel time. Liu et al.(4) developed a state space neural network 13 model for bus arrival time prediction. Chien et al. (5) used two ANN models, link based and stop 14 based, for bus arrival prediction. Park et al. (6) predicted link travel time by spectral basis 15 artificial neural network (SNN) for one through five time period ahead for same vehicle. 16 Dailey (7) used a combination of Automatic Vehicle Location (AVL) and historic 17 database to predict travel time by using KFT and statistical analysis. Cathey (8)used bus travel 18 time data as inputs to predict the same by using KFT that involved three components namely 19 tracker, filter and predictor. Shalaby (9) used a combination of AVL and Automatic Passenger 20 Counter (APC) data to predict travel time by using KFT. Nanthawichit et al. (10) used a 21 combination of Global Positioning System (GPS) equipped probe vehicles and loop detectors 22 data to estimate travel time by using KFT. The performance of the proposed methods were 23 compared with historical data based, regression and ANN models separately. 24 All the studies discussed above dealt with homogeneous traffic conditions. Only a limited 25 number of studies were reported under heterogeneous traffic conditions. Rama Krishna et al.(11) 26 used 25 trips of bus travel time to develop ANN and Multiple Linear Regression (MLR) models. 27 Vanajakshi et al. (12) used preceding two bus trips data collected by using GPS to predict next 28 bus travel time by using KFT. Padmanabhan et al. (13)extended the above study by 29 incorporating the delays in the model. Kumar(14) proposed a statistical methodology to find out 30 patterns in the data and used them as input to predict the next bus travel time by using KFT. The 31 present study will follow the above study to identify the travel time data most suited as inputs 32 and use them to develop an ANN model to predict the bus arrival time. The performance of such 33 a data driven technique will be compared with a model based approach with lower data 34 requirement (using previous two buses data). 35 DATA COLLECTION, EXTRACTION AND ANALYSIS 36 For the purpose of collecting data, GPS equipped Metropolitan Transport Corporation (MTC) 37 buses in the city of Chennai, India, were used. The test bed chosen for the present study is an 38 MTC bus route, 5C, which connects the Parry’s bus depot in the northern part of the city to the 39 Taramani bus depot in the southern part of the city. It has a route length of 15km with varying 40 land use. Figure 1 illustrates the study route with bus stop details and distances between the bus 41 stops are tabulated in Table 1. 42 TRB 2014 Annual Meeting Paper revised from original submittal. Vivek, Anil, Lelitha and Shankar 4 Table 1 Distance between Bus Stops 1 S.No Bus stop name Distance between bus stops (km) Cumulative distance from the initial bus stop (km) 1 Parry’s Corner 0.00 0.00 2 Central Railway Station 0.93 0.93 3 P. Orr & Sons 1.68 2.61 4 Wesley High School 2.70 5.31 5 C.I.T. Colony 2.34 7.65 6 Adyar Gate 2.04 9.69 7 Kotturpuram 1.96 11.65 8 Women’s Polytechnic College 2.21 13.86 9 Taramani 1.43 15.29 2