STUDY OF REDUCING THE NUMBER OF TRAFFIC SENSORS IN THE ESTIMATION OF TRAFFIC FLOWS USING GRAPH NEURAL NETWORK

• Main Author: Putri, Adiyana
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/65414
• Institution: Institut Teknologi Bandung

The increasing vehicle volume every year affects the prediction of the traffic system. The purpose of predicting traffic flow is to estimate the lost data caused by sensor malfunctions due to connection disruptions or aging. To be able to estimate the data historical from the nearest sensor is needed. The available data is modeled in the simulator to obtain data on traffic flow, speed, and road density. In the case of Central Jakarta, several sensors work, and the others are not working. To estimate data on sensors that not working, models and scenarios are needed. In this study, graph neural network models and clustering scenarios were used. The graph model through the connection of nodes with edges can describe the spatial relationship between sensors. Nodes are described as segments of sensor location close to the intersections and between nodes connected by roads or in a graph called edges. The type of graph model that used is a directed graph whose direction is determined based on the phase direction of the movement of the vehicle at the intersection. The adjacency matrix is used as a mathematical representation of the graph that state the relationship between each node, the matrix element is 1 if there is a direction that connects the nodes, while the value is 0 if there is no connection and the size of matrix is determined based on the number of nodes, then a graph representation is performed using the node2vec method and the k-means method is used to group sensor networks based on the direction of vehicle movement, the number of sensor network groups is determined from elbow method. The resulting 3 main spatial clusters in the grouping 67 sensors are divided into 9 clusters, 10 cluster and 11 clusters. Graph Neural Network (GNN) is used as a learning method that makes predictions from each sensor in the road network in each cluster. GNN gets input in the form of spatial data that has been built through a graph model in the form of an adjacency matrix and the feature data from each sensor in the form of traffic density, average vehicle speed, and delay time on the road network. The output from GNN is the prediction of traffic flow on the road network. The sensor network with 10 clusters gives an average error prediction result of 12.9%. This result is smaller than the distribution of the sensor network into 9 and 10 clusters. Each sensor cluster gradually reduces the number of sensors starting from one sub-clusters until four sub-clusters based on mean average percentage error (MAPE). With the increasing number of reduction sensors, the MAPE becomes large, the largest reduction is in the division of 10 clusters with 29 sensors reduced and the MAPE value is 38.28%. It can be concluded that even though there is a reduction in the number of sensors by clustering from a graph representation, using GNN can still estimate the calculation of vehicle flow sensors on the road network properly.