Scholarlink

<p>Investment estimation in Engineering, Procurement, and Construction (EPC) projects faces challenges such as small-sample sizes, high data noise, and complex task interdependencies. Traditional methods rely on linear assumptions, which make it difficult to effectively capture the nonlinear dependencies between the various engineering nodes of an EPC project and lack dynamic modeling of risk factors. These limitations lead to significant estimation errors and a lack of adaptability to project-specific complexities. To address these limitations, we propose Diff-PIE, a diffusion-based estimation model enhanced with transfer learning and risk-aware graph modeling. The model is pretrained on large historical datasets, fine-tuned with limited target samples, and incorporates iterative denoising for robustness. A graph-structured risk module captures risk impact factors and their propagation, enabling forward-looking assessment. Experiments on multiple EPC datasets demonstrate that Diff-PIE improves estimation accuracy by over 20% compared with existing methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Diff-PIE: an engineering investment estimation model based on transfer learning and risk-aware modeling for EPC projects

  • 期刊论文

  • 发表日期：

    2025年12月14日

查看全文
  • Judan Hu,
  • Yuyang Gao

摘要

Investment estimation in Engineering, Procurement, and Construction (EPC) projects faces challenges such as small-sample sizes, high data noise, and complex task interdependencies. Traditional methods rely on linear assumptions, which make it difficult to effectively capture the nonlinear dependencies between the various engineering nodes of an EPC project and lack dynamic modeling of risk factors. These limitations lead to significant estimation errors and a lack of adaptability to project-specific complexities. To address these limitations, we propose Diff-PIE, a diffusion-based estimation model enhanced with transfer learning and risk-aware graph modeling. The model is pretrained on large historical datasets, fine-tuned with limited target samples, and incorporates iterative denoising for robustness. A graph-structured risk module captures risk impact factors and their propagation, enabling forward-looking assessment. Experiments on multiple EPC datasets demonstrate that Diff-PIE improves estimation accuracy by over 20% compared with existing methods.