A Two-stage Model for Graph Neural Network-based Demand Forecasting and Vehicle Route Optimization in Chain Stores Delivery Planning

This paper presents a two-step predict-then-optimize framework to enhance route planning for perishable food delivery across chain stores. By addressing the challenges of demand uncertainty and geographically dispersed delivery points, we first develop a temporal forecasting algorithm based on Graph Neural Network (GNN_re) to predict multi-stores’ demand, with capturing spatial-temporal dynamics and inter-store correlations to improve forecasting accuracy. Second, a Cubic-Bernoulli mapping and Sine-Cosine Algorithm enhanced Sparrow Search Algorithm (CS_SSA) to refine routing solutions. By validated using real-world data from Sinotrans logistics company, the GNN_re achieved superior performance with an average MAE of 0.248 and RMSE of 0.148, outperforming other forecasting methods, while the CS_SSA model generated best objective value compared with other methods, meanwhile the optimized routes with predicted demand achieve minimal deviation from benchmark with actual data. This advancement provides logistics firms with actionable insights and contributes to the integration of predictive analytics in VRP research.