Real-time bus arrival prediction and bus bunching prevention using a multi-agent LSTM approach

Abstract

Bus bunching is a major challenge in urban public transportation because it reduces operational efficiency and increases passenger waiting times. Although transit agencies increasingly use real-time passenger information (RTPI) systems, existing predictive approaches often face limitations caused by sparse historical GPS data, particularly when data are collected at long polling intervals such as 10 minutes. In addition, conventional reactive control strategies frequently depend on a central dispatcher. This study addresses these limitations by proposing a scalable, proactive transit-control framework that integrates deep learning with cooperative multi-agent systems (MASs). The proposed framework uses Metropolitan Transportation Authority (MTA) Service Interface for Real-time Information (SIRI) data and introduces a forward-looking target-imputation mechanism to address data sparsity while retaining more than 90% of the original telemetry. A long short-term memory (LSTM) neural network with categorical route embeddings was developed to simultaneously predict estimated times of arrival (ETAs) at stops across New York City’s 20 highest-frequency bus routes. The predictive model was then embedded directly into autonomous bus agents within an MAS simulation. Through localized communication, the agents could cooperatively implement decentralized stop-skipping, referred to as “EXPRESS” mode, without relying on a central dispatcher. The operational effectiveness of the framework was evaluated through 120-step simulations on routes B82 and Bx9. The LSTM model achieved an overall mean absolute error (MAE) of 2.02 minutes, representing a 28.5% improvement over historical baseline models. The forward-looking imputation mechanism preserved more than 90% of the raw telemetry, demonstrating its effectiveness in mitigating data sparsity. In the MAS simulations, complete bus-bunching collapse, defined as a headway gap approaching zero, did not occur. EXPRESS mode was activated during 19.2% of simulation steps on route B82 and 25.8% on route Bx9. Following every intervention, bus headways recovered above the predefined safe threshold. The findings demonstrate that combining local deep-learning-based ETA prediction with cooperative multi-agent control can provide an effective decentralized solution to bus bunching. The proposed approach enables autonomous buses to anticipate operational disruptions and coordinate corrective actions without human intervention or centralized dispatching. Its low ETA prediction error and consistent recovery of safe headways indicate that the framework can improve both predictive accuracy and operational stability. More broadly, the study provides a scalable example of how decentralized artificial intelligence and multi-agent coordination could support proactive transit management and contribute to the development of smart-city transportation systems.