Improving public transit demand forecasting models in case of disruptions: an integrated approach using explainable AI

Abstract

Unplanned service disruptions in Public Transit (PT) systems can have major consequences on their performance and attractiveness. When such interruptions occur, PT operators must implement rapid actions to restore the service and absorb demand overflows. Among demand management strategies, short-term forecasting provides valuable information and helps to assess the levels of demand that need to be reallocated. However, short-term forecasting models struggle to take into account unexpected events such as disruptions in real-time. This work provides an integrated approach, which incorporates a Disruption Detection Module (DDM) using Random Forest together with a Demand Forecasting Module (DFM) using Long-Short Term Memory (LSTM). Results show that the integrated approach outperforms the forecasting model alone, improving prediction performance during disruption by 22%. The importance of explanatory variables is assessed with SHapley Additive exPlaination (SHAP), and the resulting analysis indicates the relevance of implementing the DDM prior to the forecasting task.