Exploring machine learning-based archetypes for urban life cycle modeling (UBiM)

Abstract

Urban analyses demand simplifications that balance modelling level of detail and scope broadness. Thus, classification by archetypes is a promising methodological approach. Such an approach is common for energy studies but rarely applied for Life Cycle Assessment (LCA) purposes. When archetypes are used in urban LCA, they generally result from previous studies for classification and characterization according to parameters that directly affect the operational energy performance of buildings. This paper tackles two research questions: i) Is it appropriate to aggregate building stocks based on operational energy (OE) variables when life cycle impacts are investigated? ii) When integrated LCA (OE + embodied impacts) is pursued, would variables describing both interests simultaneously result in better representation than using operational energy-based clustering to predict embodied impacts and vice versa? Thus, we aim to confirm that, combining variables that govern OE and embodied impacts offers a better result than using OE to predict materials groupings, even if some adherence is lost relatively to single-objective clustering. Clustering experiments were carried out for the campus of the University of Campinas, Brazil. After unsupervised k-medoid (PAM) grouping, the data were submitted to a supervised learning (neural networks) classification method. Generated confusion matrices demonstrate how adherent the clustering is when considering one interest to predict the other in three situations. Results indicate that an operational energy-driven archetype fails to represent buildings from the embodied impacts viewpoint, and that merging operational energy and embodied impact variables would better support integrated life cycle impact predictions.