Smart Traffic Prediction and Control Using Spatio-Temporal Graph Neural Networks

Project Abstract

Smart Traffic Prediction and Control Using Spatio-Temporal Graph Neural Networks presents a novel framework that integrates graph-structured representations of urban road networks with powerful spatio-temporal learning to forecast traffic dynamics and enable proactive control strategies. The study addresses the increasingly complex challenges of congestion, variability in travel demand, incidents, and heterogeneous vehicle behavior by leveraging the intrinsic topology of road networks and temporal evolution of traffic states. We propose a unified model that encodes road segments as nodes and traffic relationships (e.g., adjacent segments, turning flows, shared corridors) as edges, incorporating dynamic edge attributes such as lane closures, incidents, weather, and signal plans. The core contribution lies in designing a spatio-temporal graph neural network architecture that simultaneously captures spatial dependencies through multi-scale graph attention mechanisms and temporal patterns via recurrent and attention-based modules, enabling robust long-horizon predictions under irregular sampling and incomplete data. The methodology combines graph convolutional networks to extract spatial features with time-aware attention to model evolving traffic regimes, complemented by external data streams including probe vehicle counts, CCTV-derived occupancy, weather conditions, and event schedules. We address data sparsity and noise by employing a probabilistic fusion layer and uncertainty quantification, providing calibrated prediction intervals for incident-prone routing decisions. To translate predictive insights into actionable control, we integrate an optimization layer that optimizes adaptive signal timing, ramp metering, and route guidance in real time, subject to safety and actuator constraints. The framework supports both centralized and decentralized control paradigms, enabling scalable deployment across metropolitan and corridor-scale networks. We conduct extensive experiments on multi-city datasets featuring diverse traffic patterns, coverage densities, and roadway configurations. Evaluation metrics include mean absolute error, root mean squared error for speed and flow, and user-centered metrics such as travel time reliability and queue length reduction. Baseline comparisons span traditional time-series models, classical graph-based predictors, and contemporary spatio-temporal neural networks. Our results demonstrate superior predictive accuracy across short- to long-horizon horizons (5–60 minutes) and significant improvements in congestion mitigation when combined with the proposed control layer. The study also presents ablation analyses to quantify the contributions of graph topology, temporal modeling, and external data integration, as well as sensitivity analyses to examine robustness to data gaps, sensor outages, and atypical events. Ethical considerations include privacy-preserving data handling, bias mitigation in routing recommendations, and transparent communication of predictive uncertainty to operators. The research contributes a scalable, adaptable framework for intelligent transportation systems that marries advanced machine learning with real-time optimization, delivering improved travel efficiency, reduced emissions, and enhanced resilience of urban mobility networks.

Project Overview

What This Project Is About

---

A plain-language overview of how traffic data can be used to predict future congestion and suggest signals that adapt in real time, using a combination of spatial (where things happen) and temporal (when things happen) information with a machine learning model that understands how nearby roads influence each other.

---

---

The Problem It Addresses

---

Cities struggle with unpredictable traffic, leading to delays, pollution, and wasted fuel. Traditional models may not capture how traffic patterns spread across a network of roads or adapt quickly to changing conditions. This project aims to build a smarter predictor and controller that uses real-world road networks to improve flow and reduce stops and starts.

---

---

Objectives of the Project

---

1. Learn how traffic data is collected and stored.
2. Understand basic concepts of graphs and how roads are modeled as a network.
3. Build a simple prediction model that uses both space and time information.
4. Explore how to control signals to reduce congestion based on predictions.
5. Evaluate the model on real or simulated traffic data.

---

---

What You Will Do Step by Step

---

1. Study traffic datasets and graph representations of road networks.
2. Preprocess data (cleaning, aligning time stamps, handling missing values).
3. Implement a simple baseline model to forecast short-term traffic flow.
4. Incorporate spatial relations by modeling intersections and road links as a graph.
5. Integrate temporal patterns to capture daily and weekly trends.
6. Test different lightweight approaches before moving to complex models.
7. Simulate signal control strategies based on predictions in a small network.
8. Assess performance with metrics like accuracy and travel time improvements.

---

---

Expected Outcome

A working, explainable model that can predict near-future traffic and provide actionable signal control suggestions, alongside a simple evaluation showing potential reductions in delay and congestion for urban road networks.