Public transportation has become an indispensable part of modern life, and is used for various purposes such as commuting to work and school. Past research has shown that the convenience of public transportation has a direct impact on the willingness of users to use it. One of the ways to improve the convenience of bus routes is to provide arrival time prediction to users. However, it is very difficult to predict the arrival time accurately, especially in the case of bus routes, because of the complexity of the factors involved. In recent years, LSTM has been attracting attention for its ability to handle the characteristics of long time series, which cannot be handled by methods such as linear regression, in predicting the arrival time of bus routes.
Issues
As an extension of the LSTM-based method, a Convolutional LSTM-based method using convolutional operations for weight operations has been proposed. However, there are two major problems with this method. The first is that it does not take into account features unique to timetable-based methods, such as intentional stops at bus stops due to early arrivals, because it evaluates routes that run at regular intervals without a timetable, and the second is that its prediction accuracy drops significantly when the operation is disrupted by weather conditions such as rain. In order to solve these problems, we propose a method that simultaneously collapses weather data and past operation data using Convolutional LSTM, and uses separate models to predict travel time and stop time.
We compared the prediction accuracy of our method with that of an existing study using Convolutional LSTM for a one-week period on a bus route in Kobe City, Hyogo Prefecture, and found that the mean absolute percentage error decreased by 1.36% for the entire week. In particular, in the evening hours and during rainy weather, the mean absolute percentage error was significantly reduced by about 3% compared to the entire week.