AI-Driven Green Mobility and Traffic Optimisation Framework for Sustainable Smart Cities

Abstract

Urban transport networks are among the most significant contributors to metropolitan carbon emissions, accounting for between 23% and 31% of total city-level greenhouse gas output in major conurbations across the United Kingdom and India. As cities pursue net-zero commitments under the Paris Agreement and the UK Climate Change Act 2008 (amended 2019), the decarbonisation of urban mobility has emerged as a central operational challenge for municipal transport authorities. This paper presents AGMO-TSC (AI-Driven Green Mobility Optimisation for Traffic in Sustainable Cities), a unified deep learning framework that integrates multi-modal traffic prediction, dynamic green-wave signal coordination, intelligent electric vehicle (EV) routing and real-time carbon emission monitoring into a single cohesive urban mobility management system. AGMO-TSC employs a spatiotemporal graph attention network (STGAT) for traffic flow prediction across heterogeneous road networks, a multi-objective reinforcement learning (MORL) agent for adaptive traffic signal optimisation that simultaneously minimises travel delay and vehicular emissions and a graph neural network (GNN)-based EV routing engine that incorporates charging infrastructure availability, grid carbon intensity and real-time congestion into route selection. The framework is validated on six Indian smart city deployments — Delhi NCR, Mumbai Metropolitan Region, Bengaluru Urban Agglomeration, Hyderabad, Pune and Chennai — and three UK urban transport corridors — the M25 London Orbital, the A57/A58 Trans-Pennine route and the Birmingham City Centre Ring Road — over a 28-month operational window from March 2022 to June 2024. Across all deployments, AGMO-TSC achieves a 31.4% reduction in average intersection delay, a 27.8% reduction in transport-sector CO₂ emissions, a 23.6% improvement in EV routing efficiency and a 94.2% traffic flow prediction accuracy (MAPE 4.1%), outperforming eight baseline methods across all metrics.