Traffic Prediction in Automated Guided Vehiclular Systems using Graph Neural Networks

Abstract

Automated Guided Vehicle (AGV) systems play a critical role in modern warehouses by automating the movement of goods along set paths. This thesis investigates how to predict traffic and congestion within these systems, focusing on estimating wait times at various points in the network. We apply Graph Neural Networks (GNNs), specifically Relational Graph Convolutional Networks (RGCNs) and Relational Graph Attention Networks (RGATs), to forecast traffic conditions.

Our approach leverages a unique dataset of 50 different warehouse layouts, incorporating detailed vehicle movement simulations and traffic rules defined by fleet managers. This dataset allows us to model complex interactions and constraints affecting vehicle flow in indoor warehouses.

The results demonstrate that RGCNs effectively predict and classify wait times, achieving an F1 score of 0.94 with a quick inference time of 0.01 seconds. These findings enhance the planning and management of AGV systems by providing accurate predictions of traffic conditions, facilitating better design adjustments and reducing delays. Keywords: