Abstract:

To address complex automotive after-sales parts supply network operations with insufficient demand forecasting accuracy, slow response, and low service efficiency, this study proposed a spatiotemporal graph convolution-based method for automotive parts supply chain demand forecasting. Sales network data of the automotive parts sales network was constructed as a heterogeneous graph, integrating node features like parts sales volume and value to build multi-dimensional node dependencies. A node update mechanism of the graph convolutional neural network was designed, combined with long short-term memory neural networks to capture temporal features, using spatiotemporal attention to integrate temporal and spatial features into updated nodes. Simulation experiments show the method achieves an optimal RMSE of 2.09, error peaks within 3.44 under dynamic disturbances, and recovery within 50 steps, validating its forecasting performance and dynamic adaptability in complex supply chain environments.

Key words: spatiotemporal graph neural network, demand forecasting, parts supply chain, simulation analysis