Railway Cold Chain Freight Demand Forecasting with Graph Neural Networks: A Novel GraphARMA-GRU Model

Abstract

Accurate demand forecasting is imperative for efficient railway cold chain freight operation planning, resource optimization, and market responsiveness. Given the unique spatiotemporal characteristics and diversity of cold chain demands, the mismatch between capacity and demand has become a bottleneck, constraining the development of railway cold chain freight transportation. To tackle this challenge, we propose a graph neural network model with ARMA graph convolutional layer (ARMA Filter) and gated recurrent units (GRU), namely the GraphARMA-GRU Model, for adaptive and efficient short-term forecasting of railway cold chain freight demand. Our model can effectively capture temporal features, external factors, and the intricate spatiotemporal relationships influencing railway cold chain demands. The ARMA Filter is employed to grasp the spatial connectivity within the railway network, and GRU layers are utilized for refining temporal features. Furthermore, it also integrates external factors and refined temporal features in two graph convolutional layers to better capture multimodal characteristics. The proposed model is validated with real data of railway cold chain freight in China, whose results show an 18% improvement in prediction accuracy compared to the average performance of baseline models. In addition, interpretability methods are introduced to enhance the model’s transparency and promote future development for railway cold chain freight transportation, which may offer deep insights and support critical decisions for a smooth transition from road-based to railway-based cold chain freight transportation.