Abstract
With the increasing demand for traffic management and resource allocation in Intelligent Transportation Systems (ITS), accurate origin-destination (OD) prediction has become crucial. This article presents a novel integrated framework, effectively merging the distinctive capabilities of graph convolutional network (GCN), residual neural network (ResNet), and long short-term memory network (LSTM), hereby designated as GraphResLSTM. GraphResLSTM leverages road average speed data for OD prediction. Contrary to traditional reliance on traffic flow data, road average speed data provides richer informational dimensions, reflecting not only vehicle volume but also indirectly indicating congestion levels. We use a real-world road network to generate road average speed data and OD data through simulations in Simulation of Urban Mobility (SUMO), thereby avoiding the influence of external factors such as weather. To enhance training efficiency, we employ a method combining the entropy weight method with the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) for key road segment selection. Using this generated dataset, carefully designed comparative experiments are conducted to compare various different models and data types. The results clearly demonstrate that both the GraphResLSTM model and the road average speed data markedly outperform alternative models and data types in OD prediction.