Acta Polytechnica CTU Proceedings https://doi.org/10.14311/APP.2022.38.0169 Acta Polytechnica CTU Proceedings 38:169–175, 2022 © 2022 The Author(s). Licensed under a CC-BY 4.0 licence Published by the Czech Technical University in Prague EXPLORING MACHINE LEARNING-BASED ARCHETYPES FOR URBAN LIFE CYCLE MODELING (UBIM) Giseli Mary Colleto, Vanessa Gomes∗ University of Campinas, School of Civil Engineering and Architecture and Urbanism, Rua Saturnino de Brito, n° 224, Cidade Universitária Zeferino Vaz. CEP: 13083-889 – Campinas – São Paulo, Brazil ∗ corresponding author: vangomes@unicamp.br 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. Keywords: Archetypes, building stock aggregation, clustering, LCA, life cycle impacts, urban modelling. 1. Introduction Considering the growth projected for cities over the next decades, achieving established energy use goals become indispensable to mitigate urban environmen- tal impacts. Estimating and understanding environ- mental impacts, both operational and embodied, are essential to assist in evaluating and monitoring per- formance and to support strategic planning and the formulation of energy and environmental policies for sustainable development. From the environmental charges induced by the construction sector, studies are increasingly focusing on simultaneously reducing Operational Energy (OE) and embodied impacts [1, 2] as a decision-making ais. To avoid that only impacts during the opera- tional phase of buildings are considered [3], Life Cycle Assessment (LCA) and energy simulations must be integrated [4], to consider environmental burden over a building’s entire life cycle. Life cycle assessments at individual building scale have already been extensively covered by literature [5]. However, the need to mitigate local and global impacts to adapt to the limitations of available resources [6] makes environmental assessment at the urban scale increasingly necessary for sustainable planning. As- sessments and simulations of the urban tissue are even more complex as they face a series of challenges imposed by the scale, complexity and intensity of data needed to adequately describe urban layers and components [7]. However, if we consider the city as a set of individ- ual buildings, the high level of detail in the analysis of segregated structures requires intensive use of fi- nancial, temporal, technological and human resources. Thus, to carry out these analyses, the complexity at the urban scale allied to the lack of detailed infor- mation about the built stock demands strategies to simplify the built stock modelling, to balance level of detail and scope broadness. Thus, archetype-based approach, for instance, abstracts existing built stock into a number of reference buildings. Construction archetypes are representations (real sample or the- oretical average) of groups of buildings aggregated according to similar characteristics, defined according to the scope of the study at hand. For combining simplified modelling and extended scope, the classi- fication by archetypes is commonly used in research on new technologies [8] and energy efficiency policy design [9]. So far, their use has increased yet been focused on operational energy studies at the urban scale, but building archetypes emerge as a promising methodological approach to extend assessment scope 169 https://doi.org/10.14311/APP.2022.38.0169 https://creativecommons.org/licenses/by/4.0/ https://www.cvut.cz/en Giseli Mary Colleto, Vanessa Gomes Acta Polytechnica CTU Proceedings beyond energy use. The development of archetypes for energy studies is based on data directly related to the operational energy of buildings. According to the grouping crite- rion, the representative buildings vary, reflecting the extrapolation of characteristics to the other buildings of the group that he represents and, consequently, the consistency of the simulations for the urban scale. Specific studies assessing the environmental impact of built stock at various scales point out LCA studies ap- plying the archetype approach use the same variables as energy studies (use typology, construction typology, year/period of construction), either because they aim only at assessing operational impacts, or because they are derived from energy studies, or simply because of data availability [5]. However, an archetype for energy modeling may not be suitable for analyzing the flow of materials through the built stock [10], and representative buildings of energy perspective are not necessarily also representative from the materiality viewpoint. So, for integrated simulations of environ- mental assessments, the challenge of generating con- struction archetypes that integrate information on operational and embodied impacts arises. This paper raises 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 life cycle (OE + embodied im- pacts) assessment is pursued, would variables de- scribing both interests simultaneously result in bet- ter representation? Thus, we aim to confirm that, at the representative building identification stage of the building archetype creation process, combining variables that govern OE and embodied impacts offers a better result than using OE to predict materials groupings and vice versa, even if some adherence is lost relatively to single-objective clustering. 2. Materials and methods To achieve the above-mentioned goal, our method com- prised two main steps. Firstly, based on a systematic literature review on archetype development methods for energy studies [11], the most used clustering pa- rameters were identified according to the relevance of each study (operational energy and LCA). Then, cluster analyses to define archetypes were carried out for a selected case study, considering three different aggregation criteria: by energy-relevant variables; by materiality-relevant variables; and by combining en- ergy and materiality variables simultaneously. The case study comprised the campus of the State University of Campinas (UNICAMP), located in the municipality of Campinas, São Paulo, Brazil. With 3 893 958 m2 total area and about 50 000 daily visi- tors before the COVID19 pandemic, the study area is comparable to a small town. The campus com- prises 598 429 m2 of gross floor area (GFA) for mixed uses, such as: administration, education (classrooms), research (laboratories, workshops), health facilities (hospitals, clinics), libraries, restaurants, cultural facil- ities, sport facilities, general services, day care center, schools, bank agencies, squares, public spaces and more. For simplification’s sake, buildings smaller than the university standard building (i.e., < 500 m2 GFA) and missing data points were cutoff, so that 226 buildings remained in the final sample. Geometric and non- geometric data for detailing the sample was surveyed from several sources: GIS, field visits, image analysis, data provided by administrative sectors, and results of previous studies conducted by the research group [12]. From the operational energy standpoint, data col- lection focused on the most used parameters for EO clustering found in the literature: (i) Year of Construction; (ii) Construction Typology (isolated, semi-detached, attached); (iii) Use Typology (Administrative, Education, Hos- pital, Restaurant, Bank, Laboratory, Staff, Library, Museum, Gymnasium). From the materiality perspective, aggregation fo- cused on the structure (Reinforcing Concrete, Metallic, Structural concrete masonry, Pre-fabricated Concrete, Wood) and on the envelope (External Wall: Solid ceramic brick, Structural concrete masonry, Ceramic brick, Pre-fabricated Concrete, Wood; and Roofing: Metal, Concrete slab, Fiber cement, Ceramic, Poly- carbonate), which govern a substantial portion of building impacts, as demonstrated by life cycle as- sessment studies [1]. The Total Area and Number of Floors are relevant inputs for both perspectives. Based on a previous study on the application of different unsupervised machine learning clustering methods to the same neighborhood [7], the k-medoids / Partitioning Around Medoids (PAM) algorithm was chosen to generate a preset number of 9 clusters. In k- medoids algorithms, the groups are defined as subsets of data points that are closest to their representative elements – the medoids – which are real objects from the dataset. The identified medoids serve as the basis for developing archetypes to represent each building group. The k-medoids method is more computation- ally expensive (i.e. has longer processing time) than gravity center calculation, but is suitable if the groups are spherical, each medoid occupies a more central position in the group and can handle any kind of attribute (quantitative and qualitative). Three unsupervised learning clustering scenarios were performed: operational energy-based (OE), materiality-based (LCA) and integrated clustering (OE+LCA) (Table 1). After being processed to gen- erate clusters for each scenario and identify respec- tive medoids, the data were submitted to neural net- 170 vol. 38/2022 Exploring machine learning-based archetypes . . . Scenario Parameters used for clustering Method Cluster nr. 1 OE-related parameters: Year of Construction, Construction Typology, Use Typology, Total Area and Number of Floors 2 Materiality-related parameters: Structure, External Wall, Roof Tile, Total Area and Number of Floors 3 Integrated clustering considering simultaneously OE parameters and materiality-related parameters: Year of Construction, Construction Typology, Use Typology, Structure, External Wall, Roof Tile, Total Area and Number of Floors Unsupervised machine learning: K-medoids / Partitioning Around Medoids (PAM) 9 Table 1. Description of the clustering scenarios. (a) . Ref: OE | Pred: embodied impact. Accuracy = 45,8 %. (b) . Ref: OE | Pred: integrated LC impacts. Accuracy = 29.5 % Figure 1. Confusion matrices relating operational energy (OE) parameters (reference) to predicted embodied (A) and to integrated life cycle impacts (UBiM) (B) and respective model accuracies. works, a supervised learning classification method, using the CARET package in R language [13], and the NNET method. The statistical procedures were performed using R packages “tidyverse”, for data ma- nipulation [14]; “ggpubr”, for graph elements [15]; “cluster”, for Partitioning Around Medoids (PAM) algorithm – k-medoid [16]; “caret”, for neural net- works (supervised learning) [17] and “factoextra”, for generating cluster graphs [18]. Finally, confusion matrices were generated to demonstrate how adherent the clustering is when con- sidering one interest to predict the other in three situations: (i) considering energy grouping to predict materiality aspects (Figure 1a); (ii) considering energy to predict the integrated sce- nario (OE+LCA) (Figure 1b), and (iii) considering materiality aspects to predict the integrated scenario (Figure 2). Figure 2. Confusion matrix relating embodied im- pact parameters (reference) to predicted integrated life cycle impacts (UBiM). Model accuracy = 58.6 %. 171 Giseli Mary Colleto, Vanessa Gomes Acta Polytechnica CTU Proceedings Figure 3. Clustering based on integrated life cycle (OE and embodied) impacts parameters (scenario 3). Numbers indicate the buildings IDs. Such matrices reveal how confused the model is be- tween the predicted classes versus the actual outcomes, and highlights instances in which one class is confused for the other, providing insightful information regard- ing a model’s accuracy. The matrix columns are the “true” classes (reference), while the rows are “pre- dicted” classes. The main diagonal shows the cases where the model is correct. 3. Results and Discussion 3.1. Literature review and theoretical deepening Several efforts are being made to improve the process of creating building archetypes. De Jagger et al. [19] summarized the most common variables within 32 studies and Colleto and Silva [11] synthetized 21 publi- cations that highlighted in the archetype development process. Table 2 gathers the main parameters from those two studies, and the relevance of each variable for energy and embodied impact assessment. Other parameters have been identified as not di- rectly relevant for LCA, despite their high relevance for energy: • Building geometry: ▷ Building footprint: 4 occurrences in [19], 5 in [11]; ▷ Compactness ratio: 1 occurrence in [19], 2 in [11]; ▷ Heated volume or floor area: 2 occurrences in [19] 3 in [11]; ▷ Total loss area: 2 occurrences in [19], 1 in [11]; ▷ Density of internal thermal mass: 2 occurrences in [19]; ▷ Ground elevation: 1 case in [11]; ▷ Roof type/shape: 1 case in [19], 1 in [11]; ▷ Aspect ratio: 1 occurrence in [11]; ▷ Envelope shape: 1 occurrence in [11]; ▷ Floor loss area / Geometry / Loss-to-floor area ratio: 1 occurrence in [19] each; ▷ Position of internal thermal mass: 1 occurrence in [3]; ▷ Total area to net area ratio: 1 occurrence in [11]; • Building occupancy: ▷ Occupancy: 3 occurrences in [19]; ▷ Use of ground floor: 2 occurrences in [19]; • Thermal quality: ▷ U-value: 5 occurrences in [19], 3 in [11]; ▷ Maintenance state: 3 occurrences in [19]; ▷ Air tightness: 2 occurrences in [19]; ▷ Construction method: 1 occurrence in [19]; • Building installations: ▷ Fuel typed used: 6 occurrences in [19], 3 in [11]; ▷ HVAC system 6 occurrences in [10], 6 [11]; ▷ DHW cylinder insulation thickness / Measured energy demand: 2 occurrences in [11] each; and • Building environment: ▷ Density of urban area / Exposure: 1 occurrence in [19] each. The parameter analysis in Table 2 corroborates [10] regarding the unsuitability of an archetype for energy modeling to support materials flow analyses, by reflect- ing on the relationship of OE variables on materiality data needed for embodied impact assessment. The neural network classification and resulting confusion matrices enabled to quantify such inadequacy for the selected case study. 3.2. Sample clustering The three unsupervised clustering scenarios resulted in quite distinct results. For example, in the integrated clustering scenario 3 (Figure 3), the 9 clusters (C1 – C9) were identified in spherical groups, containing a number of elements (n): C1 = 27 n, C2 = 29 n, C3 = 10 n, C4 = 34 n, C5 = 18 n, C6 = 21 n, C7 = 39 n, 172 vol. 38/2022 Exploring machine learning-based archetypes . . . Parameters Relevance to embodied impacts Relevance to OE Bu ild in g ge om et ry Construction typology (attached, semi-detached, independent/isolated, continuous) (15) [19] (9) [11] Low – Influences the building materials proportions, but its relation to material types and masses is limited • Influences the quantities of materials (e.g., attached buildings share walls) High – It defines important characteristics for the building’s operational energy performance due to differences in solar incidence, wind, thermal mass etc. Façade-area ratio (1) [19] Roof-area ratio (1) [19] • Proportion of façade materials or roof materials out of the total amount of building materials and their respective impacts Building height (3) [19] (5) [11] Medium – Data support materials quantity surveying • Multiplier for material estimation in vertical elements (walls, columns etc) Window area (3) [19] (2) [11] • Supports window material and – jointly with WWR – wall area estimates Window-to-wall ratio (WWR) (4) [19] (3) [11] • Relative proportion of materials in windows and walls and their respective impacts Floor area (5) [11] • relevant for quantity surveying Nr. of dwellings (buildings) / rooms (residences) (2) [11] • multiplier for the bill of materials of a building floor; • materials estimation of internal elements Number of floors / stories (5) [19] (7) [11] High – Data support materials quantity surveying • Multiplier for the bill of materials for a multistorey building Total area (8) [19] • Multiplier for the total bill of materials High – Essential geometric data for operational energy performance calculations and simulations Facade materials (1) [11] • Essential data to support the bill of quantities calculations High – Non-geometric data essential for OE calculations/ simulations Bu ild in g oc cu pa nc y Use typology / Final use (14) [19] (13) [11] Low – Does not standardize building materials used in construction. Can be compensated for by correction factors that represent, for example, important differences in building systems High – equipment and use intensity/schedule are proxies for energy consumption patterns of different end use typologies. Bu ild in g ag e Construction year/period (20) [19] (19) [11] Medium – Allows retrieving the probable physical/material configuration from the implemented building regulation. In places without such regulations, it only points out building trends. High – allows retrieving the probable physical/material configuration from building regulations. If absent, utility for inferring building trends for archetypes development is limited. Year of building’s last renovation (1) [11] High – Enables potential for retrofit/renovation analyses and maintenance, which directly influence a buildings’ materiality. High – enables retrofit potential analysis Bu ild in g en vi ro nm en t Climatic zone (12) [19] (3) [11] Location (3) [19] High – Enables appropriate materials selection according to raw material sources/transportation, and energy matrix. High – Non-geometric data essential for calculations and simulations Table 2. Analysis of the main parameters used for building archetype development. Numbers within parentheses indicate the occurrences in the reviewed literature. 173 Giseli Mary Colleto, Vanessa Gomes Acta Polytechnica CTU Proceedings C8 = 48 n and C9 = 1 n. Cluster C9 contains only one building (campus hospital), for its exceptional total built area. For each cluster, the medoid (M1 = ID1367; M2 = ID1076; M3 = ID0.12; M4 = ID301; M5 = ID0.46; M6 = ID396; M7 = ID99; M8 = ID400; M9 = ID518) is the representative building to support subsequent archetype development. In confusion matrices (Figures 1 and 2), the columns are the “true” classes (reference), while the rows are “predicted” classes (in our case, building clusters). As the main diagonal shows the cases for which the model is correct, the ideal model shows a highly populated main diagonal, which sums up the highest sample percentage (in our case, 226 buildings). All other cells indicate prediction errors, so the “dirtiest” a confusion matrix is, the lowest its accuracy. Despite the accuracy achieved, applying OE pa- rameters to predict materiality grouping is the least suitable: due to the lack of data correlations, few clusters had positive results and most of them were not even statistically processed (indicated as “N/A” in Figure 1a). Contrastingly, the model’s accuracy roughly doubles if life cycle aspects prediction is based on materiality (Figure 2) instead of in OE aspects (Figure 1b): 133 of the 226 sampled buildings would be correctly grouped, and clusters C8 (41/48 buildings) and C7 (29/39 buildings) were the most accurately predicted. This reinforces our assumption that, re- gardless of its often use, using energy-based clustering and then performing LCA for embodied impacts es- timation of built stocks is not recommendable, and variables simultaneously representing both OE and materiality interests would result in better grouping than by using energy clustering for material grouping prediction and vice versa. 4. Conclusions Balancing the most appropriate parameters for cluster- ing vs. data collection and computational cost is a ma- jor challenge for creating representative archetypes to support built stock aggregation modelling. Despite improvement opportunities, the exploratory simula- tions shown support the assumption that an integrated clustering step combining operational and embodied impacts variables offers better outcomes than single- objective clustering followed by complementary simu- lation or LCA of representative buildings. Studies are underway to advance in the archetype development process and data simulation for predicting benchmarks for some use typologies and should be generally repro- ducible for varied contexts. Outcomes are expected to guide existing built stock characterization and future data collection. Acknowledgements The authors thank the National Council for Scientific and Technological Development (CNPq) for the PhD [GMC, process # 140412/2018-2] and scientific productivity [VGS, #306048/2018-3] grants. References [1] H. Birgisdottir, A. Moncaster, A. H. Wiberg, et al. IEA EBC annex 57 ‘evaluation of embodied energy and CO2eq for building construction’. Energy and Buildings 154:72–80, 2017. https://doi.org/10.1016/j.enbuild.2017.08.030. [2] L. F. Cabeza, L. Rincón, V. Vilariño, et al. Life cycle assessment (LCA) and life cycle energy analysis (LCEA) of buildings and the building sector: A review. Renewable and Sustainable Energy Reviews 29:394–416, 2014. https://doi.org/10.1016/j.rser.2013.08.037. [3] M. Österbring, É. Mata, F. Jonsson, H. Wallbaum. A methodology for spatial modelling of energy and resource use of buildings in urbanized areas. In SB14 Barcelona. 2014. http://publications.lib.chalmers. se/records/fulltext/205110/local_205110.pdf. [4] K. Allacker, V. Castellani, G. Baldinelli, et al. Energy simulation and LCA for macro-scale analysis of eco-innovations in the housing stock. The International Journal of Life Cycle Assessment 24(6):989–1008, 2019. https://doi.org/10.1007/s11367-018-1548-3. [5] A. Mastrucci, A. Marvuglia, U. Leopold, E. Benetto. Life Cycle Assessment of building stocks from urban to transnational scales: A review. Renewable and Sustainable Energy Reviews 74:316–332, 2017. https://doi.org/10.1016/j.rser.2017.02.060. [6] T. M. Baynes, T. Wiedmann. General approaches for assessing urban environmental sustainability. Current Opinion in Environmental Sustainability 4(4):458–464, 2012. Human settlements and industrial systems, https://doi.org/10.1016/j.cosust.2012.09.003. [7] V. Gomes, O. O. C. Zara, G. Colleto, et al. Archetype generation for neighbourhood lifecycle assessment. In CB2021 Parkstad. 2021. [8] C. Cerezo, J. Sokol, S. AlKhaled, et al. Comparison of four building archetype characterization methods in urban building energy modeling (UBEM): A residential case study in Kuwait City. Energy and Buildings 154:321–334, 2017. https://doi.org/10.1016/j.enbuild.2017.08.029. [9] C. S. Monteiro, A. Pina, C. Cerezo, et al. The use of multi-detail building archetypes in urban energy modelling. Energy Procedia 111:817–825, 2017. 8th International Conference on Sustainability in Energy and Buildings, SEB-16, 11-13 September 2016, Turin, Italy, https://doi.org/10.1016/j.egypro.2017.03.244. [10] International Energy Agency, Energy Conservation in Buildings and Community Systems. LCA methods for buildings. Annex 31 Energy-related environmental impact of buildings, 2001. [2022-02-16], http://www.iea-ebc.org/Data/publications/EBC_ Annex_31_LCA_Methods_for_Buildings.pdf. [11] G. M. Colleto, V. G. Silva. Review of archetype development strategies for energy assessment at urban scale. In CB2021 Parkstad. 2021. [12] O. Zara, G. Guimarães, M. Zibetti, et al. Balancing data requirement and modelling quality in neighbourhood life cycle assessments. IOP Conference Series: Earth and Environmental Science 588(4):042030, 2020. https://doi.org/10.1088/1755-1315/588/4/042030. 174 https://doi.org/10.1016/j.enbuild.2017.08.030 https://doi.org/10.1016/j.rser.2013.08.037 http://publications.lib.chalmers.se/records/fulltext/205110/local_205110.pdf http://publications.lib.chalmers.se/records/fulltext/205110/local_205110.pdf https://doi.org/10.1007/s11367-018-1548-3 https://doi.org/10.1016/j.rser.2017.02.060 https://doi.org/10.1016/j.cosust.2012.09.003 https://doi.org/10.1016/j.enbuild.2017.08.029 https://doi.org/10.1016/j.egypro.2017.03.244 http://www.iea-ebc.org/Data/publications/EBC_Annex_31_LCA_Methods_for_Buildings.pdf http://www.iea-ebc.org/Data/publications/EBC_Annex_31_LCA_Methods_for_Buildings.pdf https://doi.org/10.1088/1755-1315/588/4/042030 vol. 38/2022 Exploring machine learning-based archetypes . . . [13] R core team. R: A language and environment for statistical computing, 2021. [2022-02-16], https://www.R-project.org/. [14] H. Wickham, M. Averick, J. Bryan, et al. Welcome to the tidyverse. Journal of Open Source Software 4(43):1686, 2019. https://doi.org/10.21105/joss.01686. [15] A. Kassambara. ggpubr: ’ggplot2’ based publication ready plots. R package version 0.4.0., 2020. [2022-02-16], https://CRAN.R-project.org/package=ggpubr. [16] M. Maechler, P. Rousseeuw, A. Struyf, et al. Cluster analysis basics and extensions. R package version 2.1.2., 2021. [2022-02-16], https://CRAN.R-project.org/package=cluster. [17] M. Kuhn. Classification and regression training. R package version 6.0-90, 2021. [2022-02-16], https://CRAN.R-project.org/package=caret. [18] A. Kassambara, F. Mundt. Factoextra: Extract and visualize the results of multivariate data analyses, 2020. [2022-02-16], https://CRAN.R-project.org/package=factoextra. [19] I. De Jaeger, G. Reynders, C. Callebaut, D. Saelens. A building clustering approach for urban energy simulations. Energy and Buildings 208:109671, 2020. https://doi.org/10.1016/j.enbuild.2019.109671. 175 https://www.R-project.org/ https://doi.org/10.21105/joss.01686 https://CRAN.R-project.org/package=ggpubr https://CRAN.R-project.org/package=cluster https://CRAN.R-project.org/package=caret https://CRAN.R-project.org/package=factoextra https://doi.org/10.1016/j.enbuild.2019.109671 Acta Polytechnica CTU Proceedings 38:169–175, 2022 1 Introduction 2 Materials and methods 3 Results and Discussion 3.1 Literature review and theoretical deepening 3.2 Sample clustering 4 Conclusions Acknowledgements References