A Sageformer-KAN deep learning model for demand prediction of online food deliveries

The rapid growth of online food delivery (OFD) demand, fueled by recent consumer habit changes and technological advancements, has called for more efficient instant logistics resource management. Accurately forecasting the spatiotemporal distribution of orders to support real-time resource allocation has become an urgent challenge facing the OFD platforms. This study introduces Sageformer-Kolmogorov-Arnold network (KAN), a deep learning model for predicting food delivery demand using merchant-customer origin-destination (OD) pairs. With a multi-hop graph neural network for spatial dependencies, a transformer encoding block for temporal dependencies, and a KAN forecasting head for predictions, the model was tested on real-world data, outperforming baseline models in demand forecasting. It demonstrates excellent performance across different temporal and spatial scenarios. This study contributes to estimating the spatiotemporal order structure by forecasting future order distributions. The model aids in guiding couriers’ real-time locations, thus improving the balance between supply and demand in the OFD industry.