报告题目：Harnessing Complex Logistics Route Planning Data for Shift-Level Demand Forecasting within large-scale logistics networks

报 告 人：王明征

报告时间：2026年9月21日（星期一）10:00-12:00

报告地点：龙子湖校区06C502

主办单位：数学学院

报告人简介：

王明征，浙江大学教授、浙江大学数智仿真与智能决策技术交叉研究中心主任、浙江大学—饿了么集团数字生活联合实验室副主任、数据科学与管理工程学系副主任、中国系统工程学会物流系统工程专业委员会副主任委员、中国现代化研究会电子商务与网络空间管理分会副主任委员、中国双法研究会网络科学分会副主任委员、西藏自治区电子商务建设特聘专家、教育部大数据管理与应用专业系列教材编委会委员。主要研究兴趣包括大规模优化方法（数据驱动决策、随机优化与统计学习、大规模算法设计、大模型驱动优化方法）、数据驱动供应链与物流优化（网络设计、定价与收益管理、库存管理与物流优化、履约优化和最后一公里配送）、大数据解析与商业决策智能（人工智能可解释性需求预测方法、大模型驱动的可解释性推荐方法）等。取得的研究成果在Operations Research、M&SOM、Information Systems Research、INFORMS JOC、POM、IEEE Transactions on Automatics Control、European Journal of Operational Research, IEEE Transactions on SMC-S, Journal of Optimization Theory and Applications等国内外顶级期刊上发表70余篇，同时多篇研究成果在国际顶级学术会ICIS、INFORMS-DS和POMS ICC等上获得最佳论文奖。近几年来主持国家自然科学基金委重点项目、国家自然科学基金委重点国际合作研究项目等国家级项目十余项，撰写一步教指委指定教材《非结构数据分析与应用》。面向国家和企业的重大需求，深入淘天集团、美团集团和京东物流集团的业务场景，开展产学研合作，将产生的理论成果落地应用，解决企业运营管理难题，取得重大成效，获得了中国运筹学科学技术奖运筹学应用奖提名奖和浙江省哲学社会科学优秀成果二等奖。

报告摘要：In logistics field, shiftlevel demand forecasting within large-scale networks is highly complex due to high-dimensional and uncertain demand dynamics. Existing Graph Neural Networks (GNNs) struggle with this spatio-temporal task in volatile logistics environments due to two key limitations: a reliance on homogeneous time grids—which are inconsistent with actual asynchronous shifts—and the neglect of valuable route-specific metadata. To fully exploit GNNs, we propose restructuring temporally heterogeneous shifts into an activity network. This representation retains rich route-specific metadata and enables effective shift-level forecasting using route-planning data. To address the high volatility and execution discrepancies inherent in logistics planning, we further introduce a novel EdgeEnhanced Attention Graph (EEAG) model. EEAG effectively captures edge information, propagates spatio-temporal network effects, and mitigates the impact of input noise. We demonstrate that EEAG outperforms existing methods in both accuracy and robustness on a real, large-scale dataset from a leading logistics company, comprising over 1 million demand quantities across 2, 000+ shifts at 800+ freight transfer stations over 490 days. Compared to the best-performing baseline, EEAG reduces the forecast WMAPE by approximately 16%. This improvement facilitates effective resource allocation and route optimization, leading to over 10 million USD in annual cost savings (about 30,000 USD per day) for the company.