Thermal performance analysis of building envelope driven by attention network and decision algorithm design of energy-saving transformation

The recent advances of smart building systems have provided a significant contribution to energy prediction and optimization, although the contemporary approaches tend to be poor in terms of being able to model the intricate interactions of properties on tabular data, as well as predict and optimise individually. This work suggests a hybrid smart form that combines adaptive minimization and high feature representation for building energy management in order to address these shortcomings. The suggested approach uses a transformer-based learning model to identify correlations between numerical and categorical variables and a reinforcement learning-based method to maximize the energy system’s performance through a series of decisions. The framework is assessed using three performance metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R2). The results of the experiment show that it has high predictive performance, where the RMSE value is 2.1582, the MAE value is 1.703, and the R² value is 0.9219, indicating that it is highly accurate and robust. The efficacy of the proposed method in the context of feature interaction capturing and optimization performance enhancement is also confirmed by the comparative analysis and ablation studies.