Harnessing artificial intelligence for fleet decarbonization: a Tunisian case study using deep reinforcement learning and graph neural networks

As Tunisia strives toward a Net Zero future as a signatory of the Paris Agreement, the country is actively pursuing decarbonization strategies, by electrifying vehicle fleets to address environmental and economic challenges. This paper investigates fleet electrification through multiple transition scenarios to support decision-making. Using real-world data from a public utility fleet in Tunisia, we address two main objectives. The first objective is to model and compare three scenarios: a fully internal combustion engine fleet, a fully electric fleet, and a gradual transition scenario. These scenarios are evaluated based on their CO2 emissions and total cost of ownership. The second objective focuses on optimizing vehicle routing and scheduling to assess how operational improvements affect emissions and costs across scenarios. To support this analysis, an intelligent transportation systems decision-support framework is developed by integrating deep reinforcement learning and graph neural networks within an agent-based simulation. Results highlight the competitiveness of electric vehicles over diesel counterparts and identify key techno-economic variables from a cost-per-kilometer perspective. The transition from a fully internal combustion engine fleet to a fully electrified fleet results in a 50% reduction in CO2 emissions. After applying routing optimization, the optimized electric fleet achieves an additional 21% reduction in emissions and a 28% decrease in total cost of ownership, from USD 2.05 million to USD 1.47 million. These findings provide fleet managers and policymakers in Tunisia and other emerging economies with actionable mileage thresholds and break-even estimates to prioritize vehicle electrification and routing optimization investments.