The rapid growth in the construction sector has led to increased energy consumption and carbon emissions. Calculating energy usage and emissions is essential to energy security and promoting sustainable sector development. Therefore, the study objective is to investigate the utilazation of machine learning algorithm to predict long-term energy consumption in buildings sector, aiming to improve sustainable design and energy optimization, via the implementation of three machine learning models, XGBoost, Support Vector Regression, and Long-Short-Term Memory networks, to predict energy consumption. These models are adept at capturing complex interactions between building characteristics, environmental factors, and energy patterns. Although previous studies have explored various machine learning techniques for energy efficiency, limited research links these models to practical applications in building performance simulation. Furthermore, there is a lack of comparative evaluation of advanced machine learning models such as XGBoost, Support Vector Regression, and Long-Short-Term Memory to predict the energy consumption of building envelopes, particularly in hot climates such as the UAE. This research aims to fill this gap by providing a detailed comparison of these models against alternative approaches mentioned in the literature. The findings position Long-Short-Term Memory as a transformative force in predictive modeling, demonstrating exceptional precision with an R-squared value of 0.993 and a Mean Squared Error of 0.004. In contrast, Support Vector Regression and XGBoost showed limited predictive capabilities, with R-squared values of 0.462 and 0.94, respectively. This study establishes a solid data-driven foundation for architects and engineers to inform decisions on energy-efficient building designs, advocating Long-Short-Term Memory as the superior model for predicting energy performance.

The rapid progress of urbanization and the looming demands of 6G networks exacerbate energy consumption concerns, particularly in smart homes, where energy demand is increasingly difficult to predict and manage efficiently. Many existing studies on energy prediction in smart homes rely on limited and small-scale datasets, which restrict the robustness and generalizability of predictive models. To address these limitations, this paper introduces IntEnergy, a novel Federated Learning (FL) model designed to optimize energy efficiency in the 6G era. IntEnergy employs a regional allocation strategy where each region, consisting of multiple smart homes, assigns fog nodes to handle preprocessed data, enabling horizontal offloading between adjacent fog nodes for improved local model convergence. Utilising the portable XGBoost library in conjunction with the Flower framework, it creates a two-phase decentralized FL setup to simplify model complexity while enhancing scalability and robustness. The model leverages a comprehensive and diverse dataset to train local models in individual smart homes, ensuring improved generalizability of predictions. Experimental results demonstrate that IntEnergy surpasses state-of-the-art methods in terms of Normalized Root Mean Square Error (NRMSE) of 0.095, Mean Absolute Error (MAE) of 27.15 and R-squared value of 0.91, which highlight its efficacy in energy management in smart homes. By addressing existing limitations and introducing innovative strategies for decentralized processing and forecasting, IntEnergy provides more accurate and reliable predictions of energy consumption in the 6G era.