Development and Comparison of Deep Learning and Statistical Models to Predict Bus Passenger Flow

Abstract

To achieve long-term sustainability in the transit system, it should be periodically updated by considering demand fluctuations in the passenger flow. A prediction model with good accuracy is desirable to forecast transit demand in the future. Previous studies developed statistical and deep learning models and achieved good accuracy, but the models are restricted to that particular route or network. Hence, this study aims to discover how a prediction model behaves when it is provided with different datasets for prediction. Seasonal Auto-Regressive Integrated Moving Average (SARIMA) and Long Short-Term Memory (LSTM) models were developed for hourly and daily passenger flow prediction of one bus route, and the same models were used for the routes having different load profiles. The outcomes of the analysis demonstrate that SARIMA has the best accuracy for daily prediction compared to LSTM, but for hourly prediction, LSTM has better accuracy because of the complexity of the data. When routes with different load profiles were tested, the model's accuracy was reduced, displaying the model's incompatibility with other datasets. Hence to have a generalised model which can be adopted for diverse transit routes, the factors which affect the prediction should be identified and incorporated into the model. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.