Bus Transit Time Prediction using GPS Data with Artificial Neural Networks

Background. Bus transit time prediction is to predict how long it takes for a bus to travel from point A to point B. Good prediction helps with urban transportation planning and bus riders’ time planning. Artificial neural networks have been proven to work well in this field. Aim. This project aims to make good travel time predictions for both route and segments of route with the bus GPS data using artificial neural networks. Data. We use a collection of 925,368 latitude and longitude data points taken for buses in Porto, Portugal from April 5th to April 7th, 2015. We also used route map and bus stops data from Porto bus transportation website (http://www.stcp.pt/en/travel/) as prior knowledge. The bus GPS dataset and prior knowledge are used together to compute features for the models. Methods. We developed 3 models to make the travel time prediction, two of which are to predict whole route travel time and the other one is to predict segment of route travel time. All three models use a three-layer neural network, with different number of input and hidden units. The first model predicts the whole route travel time with features computed directly from bus GPS data and prior knowledge, the second model predicts segment of route travel time with preceding bus information and the last model predicts the whole route travel time by combining predictions of segments. Results. As the model takes more prior knowledge, higher accuracy is achieved. Moreover, whole route travel time prediction using segments has better results than the route prediction model using solely bus GPS data. It also performs better than the segment prediction model, because negative and positive errors of segment predictions cancel each other out. Conclusions. First, it is promising to develop a model that takes as many factors as possible to make accurate travel time predictions, given that prior knowledge improves model accuracy. Second, when there is lots of variations in training set, neural network may not be a good option because it tries to ”overfit” outliers. Intellectual merit. This project can serve as a basis for the development of an advanced public transportation system, which is able to make highly accurate bus transit time predictions using bus GPS data, prior knowledge and real-time traffic. Broader Impacts. With more accurate route travel time predictions, the transportation planning institute is able to arrange more efficient bus schedules and design better bus routes. As for individuals, better segment travel time predictions can provide much convenience and efficiency on time planning.