Συγκριτική ανάλυση μοντέλων υποκατάστασης μηχανικής μάθησης για την πρόβλεψη της ενεργειακής ζήτησης κτιρίων στο πλαίσιο του μοντέλου “DREEM”

Αϊβατίδη, Ελευθερία - Αγγελική

Comparative analysis of machine learning surrogate models for predicting building energy demand in the DREEM framework

Master Thesis

Author

Αϊβατίδη, Ελευθερία - Αγγελική

Date

2026

Abstract

The transition towards a more sustainable and low-carbon energy system highlights the importance of improving energy performance in the buildings sector, which remains one of the most energy-intensive sectors worldwide. In this context, Building Energy Modelling (BEM) has become an essential tool for supporting energy planning and decision-making processes. Despite their significant contribution, building energy simulation models are often associated with several challenges, including high computational requirements and limited capability for rapid adaptation to the dynamic behaviour of buildings. These limitations restrict their applicability in large-scale analyses, real-time applications, and environments requiring the evaluation of multiple scenarios. To address these challenges, Machine Learning-based surrogate modelling approaches have emerged as a promising alternative. Their main advantage lies in their ability to substantially reduce computational cost while maintaining high predictive accuracy. Within this framework, the present diploma thesis investigates and evaluates Machine Learning techniques employed as surrogate models for predicting building energy demand. Furthermore, their potential integration into the Dynamic high-Resolution dEmand-sidE Management (DREEM) model, developed by the Technoeconomics of Energy Systems laboratory (TEESlab) at the University of Piraeus, is explored. The research methodology followed is based on a structured literature review of more than 80 scientific publications, aiming to collect recent knowledge on Machine Learning and surrogate modelling approaches, as well as their development and implementation processes. In parallel, the requirements and specifications of the DREEM model were examined in order to establish an appropriate surrogate modelling framework tailored to its needs. Subsequently, a targeted analysis of scientific studies focusing on Machine Learning surrogate models for predicting building energy behaviour was conducted, followed by a comparative assessment of the identified techniques and an investigation of their suitability for integration into DREEM. The evaluation was performed using seven criteria: prediction accuracy, computational efficiency, sensitivity to training data, generalisability, robustness, interpretability, and implementation complexity. The findings indicate that no single technique consistently outperforms all others under every circumstance. Nevertheless, ensemble learning algorithms based on decision trees, particularly “Extreme Gradient Boosting (XGBoost)” and “Light Gradient Boosting Machine (LightGBM)”, emerged as the most balanced and effective solutions. These methods combine high predictive accuracy, low computational cost, and the capability to capture complex non-linear relationships. In addition, Long Short-Term Memory (LSTM) neural networks demonstrated notable advantages in applications where temporal dependencies and sequential information play a critical role. Overall, the results of this thesis suggest that “XGBoost” and “LightGBM” emerge as the most promising candidates for future implementation within the DREEM framework. However, the final selection of the most suitable surrogate model requires further experimental investigation and validation within the DREEM framework in order to confirm its performance and applicability to the specific requirements of the model.

Postgraduate Studies Programme

Βιομηχανική Διοίκηση και Τεχνολογία

Department

Σχολή Ναυτιλίας και Βιομηχανίας. Tμήμα Βιομηχανικής Διοίκησης και Tεχνολογίας

Number of pages

126

Language

Greek

URI

https://dione.lib.unipi.gr/xmlui/handle/unipi/19619

Collections

Τμήμα Βιομηχανικής Διοίκησης και Τεχνολογίας

Show full item record

Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα