Deep learning-based public transit passenger flow prediction model: integration of weather and temporal attributes

A reliable prediction model is critical for the public transit system to keep it periodically updated. However, it is a challenging task to develop a model of high precision when there is heterogeneity in the travel demand which is very common in developing countries. The spatial and temporal attributes along with external factors like weather should be incorporated into the prediction models to account for heterogeneity. Numerous studies in the past developed passenger flow prediction models considering spatial and temporal dependencies, whereas the integration of weather components with temporal dependencies while developing a prediction model for public bus transit has not been widely considered. Hence, the present research work employs long short-term memory (LSTM) to develop a route-level bus passenger flow prediction model, called RPTW-LSTM, by integrating temporal dependencies such as recent time intervals (R), daily periodicity (P) and weekly trend (T), and weather variables (W). The model is tested using a real-life dataset of the Udupi city bus service, located on the west coast of Karnataka, India. Additionally, Shapley Additive Explanation (SHAP) analysis is adopted to identify the relative importance of the features used. Results imply that the inclusion of the aforementioned factors enhanced the performance of RPTW-LSTM when compared to basic LSTM and other conventional models. Additionally, weekly trend and weather exhibit higher significance on the model than recent time intervals. This implies that evaluating the features affecting the heterogeneity in passenger flow and incorporating them into the model assists transport planners in achieving high precision.