Demand prediction model based on spatiotemporal graph neural network and external knowledge fusion

Abstract

To address the issues of insufficient spatio-temporal dependency modeling and insufficient utilization of external semantic information in the demand forecasting task in complex urban systems, this paper proposes a novel hybrid prediction model STKG-DemandNet that integrates spatio-temporal graph neural networks and external knowledge. This model innovatively combines two cutting-edge algorithms, Implicit Causal Graph Learning (ICGL) and Hypergraph Convolutional Memory Network (HCMN), which are less frequently used in demand forecasting, to construct a novel spatio-semantic joint reasoning mechanism. ICGL is used to adaptively infer potential causal driving relationships from observed data to enhance the model's sensitivity to dynamic external factors (such as sudden weather changes and emergencies); HCMN efficiently models high-order interactions among multiple entities through a hypergraph structure and introduces a memory module to capture long-term semantic evolution patterns. On this basis, the model further integrates multi-source external knowledge such as geographic information, POI semantics, and real-time event streams to achieve fine-grained and highly robust demand forecasting. Experiments on two real-world datasets (taxi order and bike-sharing usage) show that STKG-DemandNet outperforms the existing optimal methods by an average of 12.3% in MAE and 10.8% in RMSE, verifying its effectiveness and generalization ability