Carbon-aware Traffic Rerouting Using AI Foundation Models

Abstract

The global transportation sector is a major polluter, producing more than seven billion metric tons of carbon dioxide a year. The rapid increase in the number of vehicles worldwide, particularly in developing countries, continues to pose a significant challenge in managing global vehicle emissions. The spectacular growth in traffic and the resulting congestion, not only causes reduced mobility, but can also increase the emissions of the pollutants like carbon monoxide (CO), carbon dioxide (CO2), etc. Accurate traffic congestion prediction enables city planners and traffic management authorities to optimize traffic flow. However, there are several technical challenges in traffic flow prediction include unpredictable nature of traffic patterns influenced by various factors like weather, accidents, and human behaviour. Additionally, real-time data processing and the scalability of predictive models pose significant computational and infrastructure demands. To address the above mentioned challenges, we proposed a novel generalized foundation model for traffic flow prediction trained on traffic data from Caltrans Performance Measurement System (PeMS). A graph neural network has been used to represent the road network and flow/direction of traffic whereas, time series data from traffic sensors are used as a sequence of events and passed through a large language model to model the traffic sequences. Each traffic data point (e.g., a specific location at a specific time) is represented as a vector with numerical values. These values encode features like traffic volume, speed, or road type. We also derived contextual embedding to incorporate events like day of the week, accidents, holidays, etc. Further, we developed a geo-spatial foundation model to estimate aggregated emission from traffic flows and trained a model to learn the relationship between traffic volume and associated emission. This model has been used to forecast traffic emission from predicted traffic volume. Based on the predicted traffic, we developed an optimization module for adjusting traffic signal timings for minimizing the traffic emission in any smart city.