TY - JOUR
T1 - Transfer Learning for Improving Building Energy Performance and Occupant Thermal Comfort Prediction across Climates and Building Types to Address the Data Scarcity Challenge
AU - Mehraban, Mohammad H.
AU - Mirzabeigi, Shayan
AU - Ghansah, Frank Ato
AU - Sepasgozar, Samad M.E.
PY - 2026/2/28
Y1 - 2026/2/28
N2 - Accurate prediction of building energy performance is essential for energy efficiency, retrofit planning, and occupant comfort, however, traditional data-driven modeling techniques often struggle when applied across diverse climates or regions due to data scarcity and distributional differences. Transfer Learning (TL) provides a promising solution by enabling knowledge reuse from data-rich source domains to data-scarce targets and reduces the need for extensive local data collection. To investigate its potential, this study evaluates four TL techniques including instance-based, domain adaptation, parameter-based, and ensemble-based, within a hybrid stacked model that integrates a Feedforward Neural Network (FNN), Extreme Gradient Boosting (XGB), with a Gradient Boosting Machine (GBM) meta-learner. The proposed framework was trained on a high-fidelity dataset of 11,000 EnergyPlus simulations for London, which captures realistic United Kingdom (UK) residential building typologies of detached, semi-detached, attached, and apartment, with systematic variations in wall, roof, window, and floor constructions, occupancy patterns, floor area and height, window-to-wall ratio, and infiltration rate. It was then tested on 1,000 unseen scenarios from fifteen other UK cities. Evaluation of the TLs demonstrates how a London-trained AI model can be adapted to fifteen diverse UK cities with only 1% of local data points and establishes a large-scale cross-climate evaluation framework for building energy prediction under data scarcity. The zero-shot baseline achieved a mean R² of 0.718 across all climates. Instance-based and domain adaptation improved prediction accuracy in new climates by 2.6% and 10.1%, particularly in cities not having similar climate characteristics of London. Parameter-based fine-tuning pushed accuracy above 90%, while the ensemble-based reached 93%. The study demonstrates accurate, scalable, and cost-effective predictions of energy use, heating demand, and thermal comfort across diverse climates, even when the model trains mainly on a single city and then transfers to others on only 1% of target domain data. This ability to transfer models between climates and buildings provides a practical pathway for supporting energy retrofit planning, strengthening climate resilience policies across regions, and advancing affordable, clean, and energy efficient building practices in line with United Nations Sustainable Development Goal (SDG 7).
AB - Accurate prediction of building energy performance is essential for energy efficiency, retrofit planning, and occupant comfort, however, traditional data-driven modeling techniques often struggle when applied across diverse climates or regions due to data scarcity and distributional differences. Transfer Learning (TL) provides a promising solution by enabling knowledge reuse from data-rich source domains to data-scarce targets and reduces the need for extensive local data collection. To investigate its potential, this study evaluates four TL techniques including instance-based, domain adaptation, parameter-based, and ensemble-based, within a hybrid stacked model that integrates a Feedforward Neural Network (FNN), Extreme Gradient Boosting (XGB), with a Gradient Boosting Machine (GBM) meta-learner. The proposed framework was trained on a high-fidelity dataset of 11,000 EnergyPlus simulations for London, which captures realistic United Kingdom (UK) residential building typologies of detached, semi-detached, attached, and apartment, with systematic variations in wall, roof, window, and floor constructions, occupancy patterns, floor area and height, window-to-wall ratio, and infiltration rate. It was then tested on 1,000 unseen scenarios from fifteen other UK cities. Evaluation of the TLs demonstrates how a London-trained AI model can be adapted to fifteen diverse UK cities with only 1% of local data points and establishes a large-scale cross-climate evaluation framework for building energy prediction under data scarcity. The zero-shot baseline achieved a mean R² of 0.718 across all climates. Instance-based and domain adaptation improved prediction accuracy in new climates by 2.6% and 10.1%, particularly in cities not having similar climate characteristics of London. Parameter-based fine-tuning pushed accuracy above 90%, while the ensemble-based reached 93%. The study demonstrates accurate, scalable, and cost-effective predictions of energy use, heating demand, and thermal comfort across diverse climates, even when the model trains mainly on a single city and then transfers to others on only 1% of target domain data. This ability to transfer models between climates and buildings provides a practical pathway for supporting energy retrofit planning, strengthening climate resilience policies across regions, and advancing affordable, clean, and energy efficient building practices in line with United Nations Sustainable Development Goal (SDG 7).
KW - Transfer learning
KW - Building Energy Prediction
KW - Thermal Comfort
KW - Data Scarcity
KW - Cross-climateGeneralization
KW - BIM
KW - Building information modeling (BIM)
KW - Artificial Intelligence (AI)
KW - AI
KW - machine learning
U2 - 10.1016/j.enbenv.2026.02.010
DO - 10.1016/j.enbenv.2026.02.010
M3 - Article
SN - 2666-1233
JO - Energy and Built Environment
JF - Energy and Built Environment
ER -