This PhD thesis develops an innovative AI-based framework for the real-time measurement and prediction of thermal comfort, as well as for energy optimization in both indoor and outdoor environments. This research introduces the Comfort Performance Contract, grounded in the premise that energy and comfort can be accurately measure and monetized. Indoors, the framework integrates a Long Short-Term Memory network to forecast indoor temperature and relative humidity over the next 24 hours, enabling the calculation of a simplified PMV (sPMV). In parallel, a supervised learning model estimates occupants’ Thermal Sensation Vote (TSV). A Pareto-efficient optimization algorithm uses these measures to determine optimal temperature and humidity setpoints that ensure thermal comfort while minimizing energy consumption. An optimized sPMV model was also developed, combining Generalized Additive Models and regression techniques, achieving a 17.31% improvement in measurement accuracy compared to the original sPMV. An unsupervised model was also implemented, capable of identifying comfort states exclusively from physiological data, without user feedback, with a silhouette score of 0.50. For outdoor environments, an AI-based service was developed to estimate mean radiant temperature and measure the Universal Thermal Climate Index (UTCI) in real-time. This service integrates vulnerability data to enhance climate change adaptation strategies. The frameworks were validated across multiple European case studies. The LSTM model achieved mean forecasting errors of 1.1°C for temperature and 3.7% for relative humidity. The supervised TSV model obtained a mean absolute error of 0.1 (R² = 0.9), and LSTM forecasts were used to predict TSV over a 24-hour horizon. The optimization system generated recommendations capable of reducing energy consumption by up to 44.8% while ensuring continuous comfort. The outdoor MRT model achieved a MSE of 1.9°C, ensuring accurate UTCI computation.
Questa tesi di dottorato sviluppa un framework innovativo basato sull’intelligenza artificiale per la misura in tempo reale e la previsione del comfort termico, oltre che per l’ottimizzazione energetica in contesti interni ed esterni. Da questo lavoro nasce anche il Comfort Performance Contract, fondato sull’idea che energia e comfort possano essere monetizzati. Indoor, il framework introduce una rete Long Short-Term Memory che prevede temperatura e umidità interna per calcolare il PMV semplificato (sPMV) nelle 24 ore successive, mentre un modello supervisionato stima il Thermal Sensation Vote (TSV) degli occupanti. Un algoritmo di ottimizzazione Pareto-efficiente utilizza tali previsioni per fornire valori ottimali di temperatura e umidità che garantiscano comfort riducendo i consumi energetici. In parallelo è stato sviluppato un modello sPMV ottimizzato, che combina Generalized Additive Models e regressione migliorando l’accuratezza di misura del 17.31% rispetto all’sPMV originale. Inoltre, è stato implementato un modello non supervisionato capace di identificare con un silhouette score di 0.50, stati di comfort usando solo dati fisiologici, senza feedback degli utenti. Outdoor, è stato sviluppato un servizio basato su IA che stima la temperatura media radiante e calcola l’Universal Thermal Climate Index in tempo reale. Questo servizio integra dati di vulnerabilità per migliorare le strategie di adattamento al cambiamento climatico. I framework sono stati validati in diversi casi studio europei. I modelli LSTM hanno prodotto errori medi di 1.1°C e del 3.7% per il forecasting di temperatura e umidità. Il modello supervisionato di TSV ha ottenuto un MAE di 0.1 (R²=0.9). Le previsioni LSTM sono state usate per prevedere il TSV nelle 24 ore successive. Il sistema di ottimizzazione fornisce raccomandazioni che riducono i consumi energetici del 44.8% garantendo comfort continuo. Il modello MRT outdoor ha ottenuto un MSE di 1.9°C, assicurando un calcolo stabile dell’UTCI.
Development of AI-Driven Digital Services for Integrated Energy and Thermal Comfort Measurement in Indoor and Outdoor Environments under Climate Change / Cipollone, Vittoria. - (2026 Mar 19).
Development of AI-Driven Digital Services for Integrated Energy and Thermal Comfort Measurement in Indoor and Outdoor Environments under Climate Change
CIPOLLONE, Vittoria
2026-03-19
Abstract
This PhD thesis develops an innovative AI-based framework for the real-time measurement and prediction of thermal comfort, as well as for energy optimization in both indoor and outdoor environments. This research introduces the Comfort Performance Contract, grounded in the premise that energy and comfort can be accurately measure and monetized. Indoors, the framework integrates a Long Short-Term Memory network to forecast indoor temperature and relative humidity over the next 24 hours, enabling the calculation of a simplified PMV (sPMV). In parallel, a supervised learning model estimates occupants’ Thermal Sensation Vote (TSV). A Pareto-efficient optimization algorithm uses these measures to determine optimal temperature and humidity setpoints that ensure thermal comfort while minimizing energy consumption. An optimized sPMV model was also developed, combining Generalized Additive Models and regression techniques, achieving a 17.31% improvement in measurement accuracy compared to the original sPMV. An unsupervised model was also implemented, capable of identifying comfort states exclusively from physiological data, without user feedback, with a silhouette score of 0.50. For outdoor environments, an AI-based service was developed to estimate mean radiant temperature and measure the Universal Thermal Climate Index (UTCI) in real-time. This service integrates vulnerability data to enhance climate change adaptation strategies. The frameworks were validated across multiple European case studies. The LSTM model achieved mean forecasting errors of 1.1°C for temperature and 3.7% for relative humidity. The supervised TSV model obtained a mean absolute error of 0.1 (R² = 0.9), and LSTM forecasts were used to predict TSV over a 24-hour horizon. The optimization system generated recommendations capable of reducing energy consumption by up to 44.8% while ensuring continuous comfort. The outdoor MRT model achieved a MSE of 1.9°C, ensuring accurate UTCI computation.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


