Federated learning driven public transit optimization: a scalable approach for urban mobility in Addis Ababa

Purpose Urban mobility challenges in rapidly growing cities require innovative and scalable solutions to improve public transportation systems. This study aims to introduce a federated learning (FL)-based framework designed to optimize urban mobility in Addis Ababa, addressing congestions, inefficiency and privacy concerns across multiple agencies. Design/methodology/approach Using general transit feed specification data from three providers, a FL framework was used to train models locally on agency data, aggregating updates via a central server without sharing raw data. The methodology integrates long short-term memory models for arrival time prediction and reinforcement learning for route optimization. Findings The proposed methodology demonstrated substantial operational improvements, including a 25% reduction in travel time and a 20% reduction in operational costs. The optimization framework also facilitated a 30% increase in ridership, reflecting improved service reliability. Furthermore, the FL model outperformed traditional centralized approaches, improves congestion prediction accuracy by 12.5% and reduces city-wide congestion indices by 22.4%. The framework’s modular architecture supports seamless scalability, making it adaptable for diverse urban environments and larger data sets. Originality/value Addressing challenges like communication delays and data variability, this study provides a replicable, privacy-preserving model for developing cities. It demonstrates actionable insights for urban planners and policymakers, offering a replicable model for cities seeking to enhance mobility through smart and data-driven solutions. The FL-based framework contributes to sustainable urban development goals by addressing privacy concerns, optimizing routes and fostering stakeholder collaboration.