Envelope-retrofit decisions for school buildings must address overheating alongside winter energy performance. In Italy, climate classification and many associated envelope requirements are principally related to winter heating severity, while no climate-zone-specific statutory limits were identified for annual school-occupied degree-hours above 26°C. Direct overheating assessment is therefore necessary when comparing retrofit configurations across different climatic contexts. This thesis develops a machine-learning-assisted framework for predicting school-occupied degree-hours above 26°C and screening climate- and floor-specific retrofit and operational configurations. A parametric dataset comprising 46,080 cases was generated in IESVE for a school case study located in Lampedusa, Palermo, Naples, Rome, Milan and Cuneo, representing Italian Climate Zones A–F. The simulations were conducted under free-floating conditions, without active heating or cooling. The investigated variables included location, floor position, ventilation rate, orientation, window-to-floor ratio (WFR), wall and roof thermal transmittance, and predefined window packages combining U-value and solar heat gain coefficient (SHGC). Exploratory analysis revealed substantial climate- and floor-dependent variation. Lampedusa generally produced the highest overheating values and Cuneo the lowest, while Rome displayed several distinctive responses. Middle floors recorded the highest average degree-hours, although the floor hierarchy varied among climates and parameter combinations. The two investigated WFR levels, 0.12 and 0.22, produced nearly identical aggregated outcomes across most otherwise equivalent configurations. This unexpected response could not be physically explained using the available outputs. Moreover, because window U-value, SHGC and, for the existing single-glazed alternative, infiltration were coupled within predefined packages, their individual physical effects could not be isolated. Random Forest, XGBoost and CatBoost were compared using five-fold cross-validation and a separate validation subset. CatBoost achieved the strongest performance and was subsequently evaluated on the reserved test subset of 6,912 cases, obtaining an MAE of 34.384°C·h, an RMSE of 115.394°C·h and an R² of 0.997933. Most predictions closely reproduced the IESVE outputs, although substantial underpredictions occurred in a limited number of Rome top-floor cases. Permutation feature importance showed that CatBoost relied most strongly on ventilation rate, location and floor position. These rankings describe predictive reliance within the generated dataset rather than independent physical causation. In particular, the physical interpretation of the ventilation response remains uncertain because the realised airflow was not independently verified. Configuration screening also showed that no single envelope package performed best across all climate and floor groups. The results demonstrate that CatBoost can accurately reproduce and rapidly screen IESVE-generated overheating outcomes within the represented simulation domain. Nevertheless, the findings describe the thermal response of one free-floating case study and were not validated against monitored building data. The framework should therefore complement, rather than replace, detailed simulation and physical verification.
**Abstract** Le decisioni relative alla riqualificazione dell’involucro degli edifici scolastici devono considerare il surriscaldamento insieme alle prestazioni energetiche invernali. In Italia, la classificazione climatica e molti dei requisiti associati all’involucro sono principalmente legati alla severità del clima invernale ai fini del riscaldamento. Non sono stati invece individuati limiti normativi specifici per zona climatica relativi ai gradi-ora annuali oltre 26 °C durante l’occupazione scolastica. Una valutazione diretta del surriscaldamento è quindi necessaria per confrontare le configurazioni di riqualificazione in diversi contesti climatici. Questa tesi sviluppa un quadro metodologico supportato dall’apprendimento automatico per prevedere i gradi-ora oltre 26 °C durante l’occupazione scolastica e selezionare configurazioni di riqualificazione e gestione specifiche per clima e posizione del piano. È stato generato in IESVE un dataset parametrico di 46.080 casi per un edificio scolastico situato, alternativamente, a Lampedusa, Palermo, Napoli, Roma, Milano e Cuneo, in rappresentanza delle zone climatiche italiane A–F. Le simulazioni sono state condotte in regime termico libero (*free-floating*), senza riscaldamento né raffrescamento attivi. Le variabili esaminate comprendono località, posizione del piano, portata di ventilazione, orientamento, rapporto tra superficie finestrata e superficie di pavimento (WFR), trasmittanza termica di pareti e copertura, e pacchetti di serramenti predefiniti che combinano trasmittanza termica e fattore di guadagno solare (SHGC). L’analisi esplorativa ha evidenziato differenze rilevanti in funzione del clima e del piano. Lampedusa ha generalmente registrato i valori di surriscaldamento più elevati e Cuneo quelli più bassi, mentre Roma ha mostrato alcune risposte peculiari. I piani intermedi hanno presentato in media il maggior numero di gradi-ora, sebbene la gerarchia tra i piani variasse in base al clima e alla combinazione dei parametri. I due livelli di WFR esaminati, 0,12 e 0,22, hanno prodotto risultati aggregati quasi identici nella maggior parte delle configurazioni equivalenti per tutti gli altri parametri. Non è stato possibile spiegare fisicamente questa risposta inattesa sulla base degli output disponibili. Inoltre, poiché la trasmittanza delle finestre, l’SHGC e, nel caso della soluzione esistente a vetro singolo, l’infiltrazione erano associati all’interno di pacchetti predefiniti, non è stato possibile isolare i rispettivi effetti fisici. Random Forest, XGBoost e CatBoost sono stati confrontati mediante validazione incrociata a cinque partizioni e un sottoinsieme di validazione separato. CatBoost ha ottenuto le prestazioni migliori ed è stato successivamente valutato sul sottoinsieme di test riservato, composto da 6.912 casi, raggiungendo un MAE di 34,384 °C·h, un RMSE di 115,394 °C·h e un R² di 0,997933. La maggior parte delle previsioni ha riprodotto fedelmente gli output di IESVE, sebbene si siano verificate marcate sottostime in un numero limitato di casi relativi all’ultimo piano a Roma. L’analisi dell’importanza delle variabili mediante permutazione ha mostrato che CatBoost si basa soprattutto sulla portata di ventilazione, sulla località e sulla posizione del piano. Queste classifiche descrivono la rilevanza delle variabili per la previsione all’interno del dataset generato, senza dimostrare rapporti causali fisici indipendenti. In particolare, l’interpretazione fisica della risposta alla ventilazione rimane incerta, poiché le portate d’aria effettivamente realizzate non sono state verificate in modo indipendente. L’analisi comparativa delle configurazioni ha inoltre mostrato che nessun singolo pacchetto di involucro offre le prestazioni migliori per tutte le combinazioni di clima e piano. I risultati dimostrano che CatBoost può riprodurre con accuratezza e analizzare rapidamente gli esiti di surriscaldamento generati da IESVE entro il dominio di simulazione rappresentato. Tuttavia, i risultati descrivono la risposta termica di un unico caso studio in regime termico libero e non sono stati validati mediante dati di monitoraggio dell’edificio. Il quadro metodologico dovrebbe pertanto affiancare, anziché sostituire, la simulazione dettagliata e la verifica fisica.
A machine learning framework for identifying climate-adaptive envelope retrofit strategies to improve thermal comfort and reduce overheating in free-floating school buildings
YANUAR, SALMA NABILAH
2025/2026
Abstract
Envelope-retrofit decisions for school buildings must address overheating alongside winter energy performance. In Italy, climate classification and many associated envelope requirements are principally related to winter heating severity, while no climate-zone-specific statutory limits were identified for annual school-occupied degree-hours above 26°C. Direct overheating assessment is therefore necessary when comparing retrofit configurations across different climatic contexts. This thesis develops a machine-learning-assisted framework for predicting school-occupied degree-hours above 26°C and screening climate- and floor-specific retrofit and operational configurations. A parametric dataset comprising 46,080 cases was generated in IESVE for a school case study located in Lampedusa, Palermo, Naples, Rome, Milan and Cuneo, representing Italian Climate Zones A–F. The simulations were conducted under free-floating conditions, without active heating or cooling. The investigated variables included location, floor position, ventilation rate, orientation, window-to-floor ratio (WFR), wall and roof thermal transmittance, and predefined window packages combining U-value and solar heat gain coefficient (SHGC). Exploratory analysis revealed substantial climate- and floor-dependent variation. Lampedusa generally produced the highest overheating values and Cuneo the lowest, while Rome displayed several distinctive responses. Middle floors recorded the highest average degree-hours, although the floor hierarchy varied among climates and parameter combinations. The two investigated WFR levels, 0.12 and 0.22, produced nearly identical aggregated outcomes across most otherwise equivalent configurations. This unexpected response could not be physically explained using the available outputs. Moreover, because window U-value, SHGC and, for the existing single-glazed alternative, infiltration were coupled within predefined packages, their individual physical effects could not be isolated. Random Forest, XGBoost and CatBoost were compared using five-fold cross-validation and a separate validation subset. CatBoost achieved the strongest performance and was subsequently evaluated on the reserved test subset of 6,912 cases, obtaining an MAE of 34.384°C·h, an RMSE of 115.394°C·h and an R² of 0.997933. Most predictions closely reproduced the IESVE outputs, although substantial underpredictions occurred in a limited number of Rome top-floor cases. Permutation feature importance showed that CatBoost relied most strongly on ventilation rate, location and floor position. These rankings describe predictive reliance within the generated dataset rather than independent physical causation. In particular, the physical interpretation of the ventilation response remains uncertain because the realised airflow was not independently verified. Configuration screening also showed that no single envelope package performed best across all climate and floor groups. The results demonstrate that CatBoost can accurately reproduce and rapidly screen IESVE-generated overheating outcomes within the represented simulation domain. Nevertheless, the findings describe the thermal response of one free-floating case study and were not validated against monitored building data. The framework should therefore complement, rather than replace, detailed simulation and physical verification.| File | Dimensione | Formato | |
|---|---|---|---|
|
A Machine Learning Framework for Identifying Climate-Adaptive Envelope Retrofit Strategies to Improve Thermal Comfort and Reduce Overheating in Free-Floating School Buildings.pdf
accessibile in internet per tutti a partire dal 24/09/2027
Descrizione: Salma Nabilah Yanuar's Master's Thesis in Building Engineering for Sustainability
Dimensione
3.07 MB
Formato
Adobe PDF
|
3.07 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/267213