Abstract: Project material demand forecasting is a critical component of project management, playing an important role in ensuring the smooth progress of projects and reducing procurement costs. Traditional methods heavily rely on complete project information; however, in the early stages of engineering projects, information incompleteness constrains the forecasting lead time, reduces procurement flexibility, and poses significant challenges to demand planning. Moreover, project material procurement data are characterized by high dimensionality, sparsity, and heavy-tailed distributions, for which existing forecasting models exhibit limited predictive performance on extremely imbalanced datasets. To address these issues, this paper proposes a long-term project material demand forecasting method under conditions of incomplete information. By constructing an association graph structure, the method leverages complete information from similar projects to enhance prediction capabilities for early-stage projects, thereby significantly extending the forecasting lead time. In addition, to cope with data imbalance, a dedicated matching module and customized loss function are designed to improve prediction accuracy. Empirical analysis based on real project data from the State Grid Corporation of China spanning 2015–2023 demonstrates that transferring source-domain information from projects with complete information to early-stage projects with incomplete information can effectively enhance forecasting performance. Furthermore, when detailed project information is lacking in the early stages, standardized and fine-grained material coding schemes play a particularly important role in demand forecasting, with the precision contribution of a single indicator exceeding 20%. The proposed method not only achieves strong predictive accuracy in the early stages of projects but also outperforms traditional demand forecasting models overall. This paper provides an effective tool for project material demand forecasting and contributes to the enhancement of digital and intelligent operational management of engineering projects.

Key words: demand forecasting, project materials, incomplete information, correlation matching, graph neural network model

摘要： 工程物资需求预测是工程管理中的关键环节，对保障工程顺利推进、降低采购成本具有重要意义。传统的工程物资需求预测依赖于工程项目的完全信息，但在工程初期，信息的不完全制约了需求预测的提前期，降低了采购执行裕度，给需求计划制定带来挑战。此外，工程物资采购数据具有高维、稀疏和长尾特性，现有模型对这类极度不平衡数据集的预测效果有限。针对这些问题，本研究提出了一种信息不完全条件下的工程物资需求远期预测方法。该方法通过构建关联图结构，利用相似工程的完全信息增强对早期项目的预测能力，显著提升了需求预测的提前期；同时，针对数据集分布不平衡问题，设计了专用匹配模块与损失函数，提高了预测精度。基于国家电网2015~2023年真实工程数据的实证研究表明：将完备信息项目的源域信息迁移至信息不完备的初期项目，可有效提升预测能力；在工程早期缺乏详尽工程信息的情况下，规范、精细的物资编码方式对物资需求预测尤为重要，单一指标的精度贡献占比可达20%以上；本方法不仅在工程初期表现出良好的需求预测精度，而且整体优于传统需求预测模型。本研究为工程物资需求预测提供了有效工具，有助于提升工程项目的数智化运营管理水平。

关键词: 需求预测, 工程物资, 不完全信息, 关联匹配, 图网络模型