Accurate and actionable pallet demand forecasting is essential for multi-echelon warehouse supply chains, particularly when facing complex inter-node dependencies and time-varying transportation lags. Existing methods, despite progress in modeling supply chain dynamics, still struggle with temporal misalignment of multi-source signals and with enforcing operational constraints, limiting the direct usability of their forecasts. We propose SC-DGLA, a constraint-aware forecasting framework that integrates dynamic graph learning with conditional learnable lag alignment (LLA). It employs dynamic graph modules to capture evolving network structures and perform multi-task edge-level predictions, and incorporates conditional LLA to temporally align production, transfer, sales, and return signals. Constraint-aware training with projection-based decoding then ensures feasibility and yields decision-ready outputs. Experiments on real-world pallet data from the central warehouse of a large retail enterprise show that SC-DGLA maintains accuracy comparable to strong baselines while achieving a forecast feasibility rate of 92.8% and reducing shortage/overcapacity rates to 3.1% and 2.7%, respectively, offering more practical and operationally feasible forecasting support for warehouse planning and decision-making.
Ye et al. (Fri,) studied this question.