Machine learning based travel time prediction of urban bus transit using GPS data

Abstract

An accurate and reliable arrival time prediction of buses to the next bus stops is a valuable tool for both passengers and operators. Existing studies have some limitations in bus travel time prediction. They focus little on three aspects such as heterogeneous traffic flow conditions, dwell time prediction and interpretation of explanatory variables. Consequently, we break down the prediction problem into sub-models for running times and dwell time prediction and incorporate a feature engineering framework that generates features related to the running bus, the prediction day, and immediate and long historical time variations to capture heterogeneous traffic conditions. We propose a multi-model stacked generalisation ensemble model by leveraging the advantages of best-performing models in homogeneous conditions such as Extreme Gradient Boosting (XGBoost) and convolutional long short-term memory (ConvLSTM) models. It outperformed the state-of-the-art models by 11% in mean absolute error (MAE) on average. It can predict extreme conditions in bus journeys more accurately. Nevertheless, the input data for the machine learning model should be the historical travel times of the route. We proposed two simple novel algorithms to extract bus trips and match bus stop sequences towards extracting dwell times and running times from the raw crude GPS data generated at a medium sampling frequency of 15 seconds. Those algorithms incorporate various challenges like non-uniformity, poor network coverage, discontinuities in streaming and skipping of bus stops. In addition, we attempted to interpret the feature importance of the generated features. We found insights like driver behaviour and the immediately preceding dwell time influence the stopping pattern and the prediction model, which pave the way for strategic management by authoritie