Spatio-temporal information enhance graph convolutional networks: A deep learning framework for ride-hailing demand prediction.

Abstract

Ride-hailing demand prediction is essential in fundamental research areas such as optimizing vehicle scheduling, improving service quality, and reducing urban traffic pressure. Therefore, achieving accurate and timely demand prediction is crucial. To solve the problems of inaccurate prediction results and difficulty in capturing the influence of external spatiotemporal factors in demand prediction of previous methods, this paper proposes a demand prediction model named as the spatiotemporal information enhance graph convolution network. Through correlation analysis, the model extracts the primary correlation information between external spatiotemporal factors and demand and encodes them to form feature units of the area. We utilize gated recurrent units and graph convolutional networks to capture the spatiotemporal dependencies between demand and external factors, respectively, thereby enhancing the model's perceptiveness to external spatiotemporal factors. To verify the model's validity, we conducted comparative and portability experiments on a relevant dataset of Chengdu City. The experimental results show that the model's prediction is better than the baseline model when incorporating external factors, and the errors are very close under different experimental areas. This result highlights the importance of external spatiotemporal factors for model performance enhancement. Also, it demonstrates the robustness of the model in different environments, providing excellent performance and broad application potential for ride-hailing prediction studies.