Deep Learning-Driven Contextual Stochastic Optimization for Real-Time Order Fulfillment

Order fulfillment optimization is a fundamental challenge in large-scale e-commerce, requiring real-time decisions for every incoming order. For enterprises with extensive fulfillment networks, selecting the optimal fulfillment plan demands balancing operational costs with strict service-level guarantees under uncertainty. To model this problem, this work introduces a two-stage contextual stochastic optimization framework explicitly capturing two sources of uncertainty, delivery timeliness and future inventory consumption. To enable real-time deployment in peak hours, where traditional solvers are computationally prohibitive, an optimization proxy is developed, training deep neural networks to approximate solutions of the underlying stochastic program with high fidelity. Computational experiments on a large-scale JD.com transactional dataset demonstrate that the proposed approach achieves orders-of-magnitude speedups compared to a state-of-the-art commercial solver while preserving similar solution quality. The results establish a scalable paradigm for real-time stochastic optimization in e-commerce logistics, bridging rigorous optimization with deep learning to deliver industrial-scale efficiency