| description abstract | The postpandemic period has seen a significant increase in traffic volume on freeways, necessitating the implementation of advanced traffic management systems, such as lane-level freeway tolling systems, to predict traffic patterns and alleviate congestion. Although deep learning models have proven effective in predicting traffic states, little research has focused on lane-level traffic prediction, which is crucial for emerging intelligent transportation applications. To address this gap, this study develops a lane-level road segment graph and proposes a lane-based road network traffic volume prediction model, GCN-LSTM, that combines graph convolution network (GCN) and long short-term memory (LSTM). The proposed model employs different graph Laplacian matrices, and the performance of these corresponding derived models is compared with that of existing traffic prediction models. The proposed model is evaluated using traffic volume data collected from inductive loop detectors installed on freeways in the Seattle area, including both high-occupancy toll lanes and general-purpose lanes. The results demonstrate that the GCN-LSTM model with the combinatorial Laplacian matrix outperforms other models. Additionally, the model’s prediction performance remains consistent when using input data with various temporal ranges. Furthermore, excluding high-occupancy toll lane data from the dataset improves the prediction accuracy, highlighting the importance of developing specialized models for lane-level traffic prediction tasks. | |
