supply chain

In supply chain decision-making, optimization recommendations suffer from poor interpretability, complex human–system interaction, and delayed model updates—resulting in decision cycles spanning days to weeks and heavy reliance on data science teams. This paper introduces the first LLM-powered intelligent interaction layer tailored for supply chain optimization, integrating natural language understanding, knowledge-based reasoning, and optimization tool orchestration. Our approach delivers three key contributions: (1) automated generation of human-interpretable explanations for optimization outputs; (2) natural-language-driven dynamic scenario simulation and “what-if” analysis; and (3) business-feedback-guided adaptive retraining and updating of mathematical optimization models. By eliminating manual intermediation, the framework reduces decision latency to minutes, significantly enhancing autonomous decision-making capabilities for planners and executives. It advances the democratization and real-time operation of supply chain decision technologies.

Retail demand data are often plagued by strong seasonality, irregular spikes, and noise, which undermine the accuracy of traditional forecasting methods and hinder effective supply chain decision-making. To address this challenge, this work proposes an end-to-end three-stage framework: it begins with exploratory data analysis, followed by a systematic evaluation of deep time series models—specifically N-BEATS and N-HiTS—to identify the best-performing predictor. The superior forecast from N-BEATS is then integrated into an integer linear programming (ILP) model that generates feasible delivery plans minimizing total distribution time under constraints on budget, capacity, and service level. By combining high-accuracy deep learning forecasts with interpretable constrained optimization, the approach successfully translates four-week demand predictions for 1,918 units into cost-optimal, executable logistics plans, substantially enhancing operational efficiency.

This study addresses the challenge of integrated logistics and production scheduling in flexible, personalized pharmaceutical manufacturing by proposing a unified optimization framework that simultaneously considers bin packing, equipment layout, task scheduling, and path planning for automated production lines based on planar transport systems. The approach leverages drug co-occurrence patterns and Hamiltonian path-based neighborhood optimization to determine equipment placement, formulates bin packing and layout as a mixed-integer quadratic program, employs constraint programming for task scheduling, and generates conflict-free vehicle routes through directed acyclic graph reasoning coupled with iterative conflict resolution. Experimental results on 40 real-world prescriptions demonstrate that the system efficiently handles up to 500 daily orders across diverse facility layouts, achieving high performance while maintaining computational tractability.

To address fragmented forecasting strategies, weak data-driven capabilities, and the absence of decision闭环 in supply chain forecasting, this paper proposes a KPI-driven big-data prediction management framework. Methodologically, it integrates multi-source heterogeneous data collection, phantom inventory impact modeling, and hierarchical periodic forecasting strategies, coupled with XGBoost/LSTM ensembles, Bayesian hyperparameter optimization, and dynamic preprocessing. This enables a closed-loop workflow spanning problem identification, modeling, and feedback. The key contribution is a novel KPI-oriented paradigm that tightly couples preprocessing, prediction, and decision-making—marking the first integration of inventory, workforce, and capacity KPIs into both feature engineering and feedback-driven model refinement. Empirical results demonstrate an 18.3% average reduction in MAPE for mid-to-long-term demand forecasting, a 22% improvement in inventory turnover ratio, and a 35% reduction in planning response cycle time, significantly enhancing forecast transparency and operational decision agility.

This paper addresses the last-truck scheduling problem in e-commerce middle-mile transportation, aiming to maximize next-day delivery order fulfillment under fixed inventory locations and sufficient last-mile delivery capacity. We formulate this problem as an NP-hard submodular optimization problem with coverage constraints—the first such formulation in the literature. To solve it, we propose three scenario-adaptive algorithms integrating greedy submodular maximization, pipage rounding, and Lagrangian relaxation heuristics, achieving both theoretical guarantees (worst-case approximation bounds) and scalability. Extensive experiments on real-world logistics networks and datasets demonstrate that our solutions closely approximate the global optimum, significantly improving next-day delivery coverage while maintaining computational efficiency. The proposed framework is production-ready and demonstrates strong industrial deployability.

本文提出一种混合代理AI框架，通过协调代理解析用户意图并分配任务给专门代理，解决供应链分析中的决策难题，提高效率和成本效益。

研究通过结合图神经网络、变邻域搜索和集合划分重组的方法，优化了两层备件网络设计，解决了高成本评估问题，并提高了节约率。

This study addresses the inefficiencies in traditional supply chains, where isolated optimization of individual decision modules often leads to inventory imbalances and elevated transportation costs. To overcome this limitation, the authors propose an end-to-end collaborative composite policy model that, for the first time, formulates multi-stage operational decisions—product assortment, sourcing allocation, replenishment frequency, and delivery routing—as a sequentially coupled token generation process. Supply chain entities are represented as tokens, and decisions are generated in sequence through a shared contextual representation and a serialized decision interface, enabling tight coupling across stages. The entire replenishment plan is evaluated holistically via a system-level utility function. Experiments on real-world datasets from Dingdong Maicai and JD.com demonstrate that the proposed approach significantly outperforms both stage-wise independent optimization and industry baselines, confirming the effectiveness of cross-module joint learning in enhancing overall supply chain performance.

This study addresses three critical challenges in e-commerce supply chain planning—model fragmentation, computational non-scalability, and low decision credibility—by proposing the SPORD framework. SPORD introduces a novel decoupled simulation-and-optimization architecture that transforms simulation from a monitoring tool into a planning engine: feasible logistics paths are first generated via simulation and then an integer programming model selects the globally optimal subset, enabling end-to-end collaborative planning. The method integrates matrix-vectorized CPU/GPU-accelerated simulation, list scheduling, and integer programming to efficiently solve large-scale instances with numerous SKUs and network nodes. Furthermore, a closed-loop intelligent diagnostic mechanism enhances solution interpretability and reusability. Since its deployment in 2025, SPORD has served over 20,000 suppliers, reducing cross-regional fulfillment rates from 6.1% to 4.9% and achieving an average monthly carbon reduction of approximately 5,745 metric tons of CO₂ equivalent.

研究通过结合大数据和大型语言模型自动化设计库存策略，使用迭代生成和优化参数的方法，有效降低了库存成本并发现新的策略形式。