A Meta-Learning Approach for Predicting Embodied Carbon Emissions for Ready-Mix Concrete Products

초록

In recent years, uncertainties in carbon estimation models have posed significant challenges for stakeholders in identifying embodied carbon hotspots, particularly amid increasing demands for standardized assessment methodologies across jurisdictions. To address this issue, this study employed a Model-Agnostic Meta-Learning (MAML) algorithm to predict carbon emissions using 12,000 Environmental Product Declarations (EPDs) documents of concrete mixtures from North America. The MAML model was trained to predict embodied carbon for ready-mix concrete using resource use, material consumption, and waste generation data from the United States (U.S.) as base learners and was evaluated on a new prediction task in Canada. The results demonstrated that the MAML model exhibited strong generalization performance, achieving an R2 score of 0.947, outperforming the base learners, which recorded R2 scores of 0.807, 0.772, and 0.684. Additionally, reductions in RMSE and MAE ranged from 29% to 33%, underscoring the model’s effectiveness in handling few-shot learning scenarios with previously unseen features. These findings suggest that the embodied carbon prediction model based on the MAML algorithm can be generalized across jurisdictions and support policymaking toward sustainable development in the construction industry.

키워드