Trafikruttplanering med hjälp av grafiska neurala nätverk

Sammanfattning: With urbanization and cities becoming more congested, the need for effective traffic flow management is increasing. With real-time traffic sensors, the current traffic flow can be measured and more effective transportation routes in cities can be made using this information. Using the data from traffic sensors, a graph neural network (GNN) can be trained to predict future traffic flow. GNNs are neural networks specifically made to handle graph-structured data, which road networks can be represented by. Synthetic traffic flow data at different intersections were used to train the model. The data was collected using the \textit{simulation software Simulation of Urban Mobility}, which is widely used in academic literature. The data is structured in one-minute intervals, where the model was trained to predict one minute into the future. Test scenarios were constructed where a vehicle had to find a route between two nodes, inside the simulated road network. This was done to test the effectiveness of the sensors and model. The tests were carried out by seven digital twins, one baseline that used the fastest path, three that used the real-time sensor data, and the last three that used the model to predict future traffic flow. It was found that the digital twins that used both sensor data and the trained GNN model to find the route performed on average better than the baseline in the metrics: time traveled, Co2 emission, time stopped, and average speed.