Development of Surrogate Models for Estimation of Energy Consumption, Carbon Emissions, and Thermal Comfort in a Residential Building

Estimating energy consumption and CO 2 emissions in buildings is crucial for mitigating global warming. This paper outlines the development of surrogate models aimed at estimating key output variables, namely annual energy consumption, annual CO 2 emissions, and predicted percentage of dissatisfied. The surrogate models are constructed using three machine learning (ML) methods: decision trees (DTs), group method of data handling (GMDH), and random forest (RF), as well as a deep learning (DL) approach, feedforward neural networks (FNNs). Their accuracy is rigorously evaluated using assessment metrics such as root mean square error (RMSE), mean absolute percentage error (MAPE), and R‐squared ( R 2 ), complemented by computational time analysis. These are the main contributions of the present paper. The results indicate that FNN achieves the highest accuracy across all performance metrics, with an R 2 value of 0.999 and the lowest prediction errors. In contrast, DT exhibits the least accuracy but requires the shortest computational time. Additionally, an analysis of the impact of design variables on energy consumption and thermal comfort reveals that the heating setpoint (HSP) temperature has the most significant effect, while building orientation has the least. These findings highlight the potential of DL–based surrogate models for optimizing building energy efficiency. The prospect of integrating DL–based surrogate models with optimization algorithms emerges as a promising avenue for effectively optimizing energy consumption in buildings.