Dynamic Heterogeneous Graph Neural Network with Carbon-Sensitive Dual Attention for Lifecycle Carbon Footprint Assessment of Engineering Projects

Abstract

Global engineering projects are major carbon emitters with high heterogeneity, but traditional assessment methods (e.g., LCA, IPCC) lack precision, efficiency, and adaptability to dynamic construction. This study proposes a Carbon Footprint-aware Graph Neural Network (CF-GNN) for lifecycle carbon assessment. Its core innovations include: (1) a dynamic heterogeneous graph (entity/attribute nodes) updated via 15-day cycles and milestone triggers; (2) a carbon-sensitive dual attention mechanism prioritizing high-emission nodes/edges; (3) a third-order message passing framework capturing multi-hop carbon flows (up to 5 nodes). Validated on 3.86 million time-series data from 16 projects (residential, bridge, factory, etc.) against 8 baselines (LCA, GAT, TGAT, etc.), results show: CF-GNN achieves an average MAPE of 7.2% (38.9% lower than GAT, 55.8% lower than LCA), with bridge project RMSE at 218 tCO₂ (59.4% lower than LCA). It has 2.0±0.1s inference latency for 1000 nodes and 52±3.1min end-to-end assessment—3375-fold less manual effort than LCA (6 months/bridge). Key node identification matches experts (0.87 Kendall coefficient), with CV<5% (high stability) and 94.2±1.5% coverage for 95% prediction intervals. CF-GNN enables precise, efficient dynamic assessment, supporting low-carbon design/optimization and advancing "dual carbon" goals in construction, transportation, and energy.