In Sect 4, to validate the effectiveness of the proposed method, we selected several representative datasets, including the “BIM-ECA Dataset” [43], “NEBULA Dataset” [44], “BEC Dataset” [45], and “BIM-BEM Dataset” [46], to ensure the broad applicability and reliability of the experimental results. The experimental design covers various aspects, from data preprocessing, feature extraction, and model training to performance evaluation, with a focus on examining the comprehensive performance of Graph Neural Networks, Transformer models, and Generative Adversarial Networks in building energy efficiency optimization. The schematic diagram of the building energy efficiency optimization experiment is shown in Fig 5.
4.1 Implementation details
4.3 Evaluation index
In this study, we develop a multi-metric evaluation system to assess the effectiveness of building energy efficiency optimization. Energy consumption, the basic metric, is recorded over a specific period and provides a preliminary assessment of the optimization strategy’s energy-saving effect. The energy saving rate, which represents the percentage change in energy consumption before and after optimization, intuitively demonstrates the contribution of energy-saving measures. The energy utilization rate examines the efficiency of energy conversion from input to useful output in building systems, thereby revealing energy losses. These complementary metrics form a comprehensive framework for analyzing the effectiveness and value of the optimization strategy. The following sections define the calculation methods and applications of these metrics to ensure a scientific and thorough evaluation of the optimization outcomes.
- Energy Consumption
Energy consumption quantifies the total energy used by a building over a specific period, serving as a fundamental indicator in energy efficiency assessment. This study employs Graph Neural Networks, Transformer models, and Generative Adversarial Networks to simulate and predict building energy use, providing a scientific basis for optimization. Energy consumption includes all forms (electricity, heat) utilized by a building, measured in kWh or MWh. The total building energy consumption is calculated as:
(8)where Etotal represents total consumption over a period, P(t) is power demand at time t (kW), and 
is the time interval (hours).
In practical calculations, P(t) is influenced by equipment operation, climate conditions, and building usage. For precise assessment, we refine the power demand calculation:
(9)where Pbase(t) represents baseline consumption (thermal load), PAC(t) is air conditioning demand, Plighting(t) is lighting demand, and Pequipment(t) covers other equipment consumption.
Beyond direct consumption calculations, we employ derived indicators including energy saving rate and energy utilization rate to comprehensively assess building performance across usage stages and renovation schemes. This approach evaluates both total consumption and efficiency improvements from various optimization strategies.
Energy consumption thus serves as a core metric for building efficiency evaluation, dependent on operational data and environmental factors. Through our advanced methods, we accurately predict and optimize building energy use, supporting effective energy management and carbon emission reduction.
The energy saving rate (ESR) quantifies building energy efficiency improvements by comparing consumption before and after upgrades. It demonstrates the effectiveness of green renovations and efficiency enhancements through the following formula:
(10)Here, Ebaseline represents original energy consumption before implementing energy-saving measures, while Eoptimized is consumption after optimization. The percentage result indicates energy saved after efficiency optimization, with higher ESR values reflecting more successful interventions.
In practice, Ebaseline derives from historical building data under normal operating conditions (heating, cooling, lighting, etc.), collected under consistent environmental and operational parameters to ensure evaluation accuracy. Eoptimized represents consumption after implementing efficiency measures through green renovations, energy-saving technologies, and system optimizations. This value is determined through experiments or simulations that combine operational data with environmental conditions.
In our study, ESR calculation extends to different optimization schemes. Using the GNN model to predict building spatial structure and energy efficiency, we simulate how various structures impact energy consumption, guiding green renovation plans. ESR improvements typically result from design-phase optimization, construction-phase energy-saving measures, and operational-phase efficiency management. We adjust optimization strategies based on building types, environments, and budget constraints.
The energy utilization rate (EUR) measures effective energy use in buildings and reflects resource efficiency. As a fundamental metric in optimization research, EUR quantifies the practical returns of energy input. The formula is:
(11)In practical applications, Euseful represents energy used for specific building functions (heating, lighting, etc.). System efficiency varies with equipment status and environmental conditions. Our study uses Graph Neural Networks to analyze relationships between building structures and energy demand, optimizing distribution and improving utilization efficiency. The Transformer model captures temporal patterns of energy consumption, enhancing prediction accuracy and enabling precise energy requirement determination during energy-saving plan design.
Conversely, Etotal encompasses all energy consumption, including system losses and transmission inefficiencies. Heat loss through building envelopes and other inefficiencies contribute to total consumption without serving building functions. Accurate measurement of these factors ensures comprehensive evaluation of utilization rates.
Improving EUR requires optimizing building design, equipment operation, and energy management. High-efficiency systems, improved insulation, and intelligent control reduce energy losses. In our study, GAN-generated renovation plans simulate different energy-saving schemes’ impacts on utilization efficiency. Optimized facades, energy-efficient materials, and intelligent regulation systems boost utilization rates, reducing total energy use and greenhouse gas emissions while advancing the construction industry’s green transformation.
In summary, EUR represents a critical indicator in building energy efficiency optimization, reflecting actual consumption efficiency. Our study combines advanced deep learning methods to optimize building design, operation, and energy management, contributing to sustainable development in the construction industry.
We employ three standard regression metrics to quantify prediction accuracy. The Root Mean Squared Error (RMSE) is defined as:
The Mean Absolute Error (MAE) is calculated as:
The Coefficient of Determination (R2) measures the proportion of variance explained:
where 
and 
represent predicted and actual energy consumption, N is the number of test samples, and 
is the mean of actual values.
4.4 Experimental comparison and analysis
In the previous sections, we introduced the building energy efficiency optimization methods proposed in this study. These intelligent approaches aim to enhance energy efficiency and reduce consumption by modeling and optimizing building energy use. In the experimental part, we analyzed and compared different model architectures and optimization strategies to verify their effectiveness in improving building energy efficiency.
During the comparison and analysis, we focused on evaluating the performance of different methods in terms of prediction accuracy, computational efficiency, and model stability, while also considering the adaptability and generalization ability of different models across various datasets. Through these comparative analyses, we aim to provide clearer directions for future building energy efficiency optimization research and offer more scientific theoretical support for the green and low-carbon transformation of the construction industry.
In this study, we integrated BIM data with machine learning techniques to conduct comprehensive analysis and visualization of building energy consumption. Initially, we utilized the BIM model to collect detailed energy consumption data from various functional areas of the building. This data encompassed key energy-consuming components including heating, ventilation, hot water supply, fans, pumps, lighting, technical equipment, elevators, cooling, and ventilation cooling.
Following data analysis, we generated two line charts illustrating the annual energy consumption (kWh/year) and annual energy consumption per unit area (kWh/m2·year) for each functional area. These visualizations clearly demonstrated that hot water supply and ventilation cooling constitute the primary energy consumption sources, while technical equipment and lighting also contribute significantly to the building’s overall energy usage.
This visualization analysis enabled us to identify critical energy consumption hotspots and establish clear directions for subsequent energy efficiency optimization. By incorporating machine learning models, we uncovered underlying patterns within the energy consumption data. Specifically, we employed the self-attention mechanism to focus on key energy factors and utilized graph neural networks to capture complex relationships between various building components. Furthermore, Generative Adversarial Networks generated multiple optimization solutions based on this data, providing scientific evidence to support green renovation strategies.
Through this data-driven approach, we developed more precise energy-saving strategies, thereby promoting improved building energy efficiency and sustainable development. The energy consumption results are shown in Fig 6 below:
In the energy efficiency evaluation of Tables 1 and 2, the proposed method demonstrates significant performance advantages. On the BIM-ECA dataset, compared to the method of Chen et al., the energy saving rate increased by 4.75 percentage points, and the energy utilization rate increased by 4.59 percentage points; compared to the method of Pan et al., the increases were 3.2 and 3.37 percentage points, respectively. The results on the BEC dataset are particularly outstanding, with the energy saving rate improving by 0.86 percentage points and the energy utilization rate increasing by 2.69 percentage points compared to the method of Dounas et al. Notably, based on the method of Leite et al., the proposed method shows small but significant improvements in both energy saving rate and energy utilization rate. The results on the NEBULA and BIM-BEM datasets also indicate that the proposed method consistently maintains a leading position in energy efficiency indicators, excelling not only in individual metrics but also demonstrating stability and applicability across multiple datasets. This consistency across datasets suggests that the proposed method has strong adaptability and scalability, effectively enhancing the overall performance of building energy management systems. Through comparative analysis, we can clearly observe the significant progress of the proposed method in energy saving and utilization efficiency, providing new insights and solutions for the field of building energy management. The experimental results of Tables 1 and 2 are visualized in the following Fig 7:
Based on the experimental results in Tables 3 and 4, it is clear that the proposed method demonstrates exceptional advantages in computational performance metrics, significantly outperforming other research methods in terms of training time, inference time, and model parameter count. On the BIM-ECA dataset, compared to the method of Chen et al., the training time was reduced by approximately 34.07%, inference time decreased by 17.7%, and the model parameter count was reduced by 22.5%. Compared to the method of Leite et al., the training time was compressed by 6.09%, inference time decreased by 2.54%, and the parameter count dropped by 3.93%. On the NEBULA, BEC, and BIM-BEM datasets, the proposed method also maintains a significant performance advantage, with training times generally under 35 seconds, inference times controlled within 135 milliseconds, and model parameter counts kept below 2.45 million. This consistency across datasets indicates that the proposed method not only possesses efficient computational performance but also demonstrates good scalability and generalizability. Through comparative analysis, it can be seen that the proposed method significantly reduces computational resource consumption while maintaining high energy efficiency, providing a more economical and efficient technical solution for building energy management. This has important practical implications for promoting green buildings and energy conservation. Similarly, we have visualized the experimental results from Tables 3 and 4, as shown in the following Fig 8:
Table 5 summarizes the evaluation of the GNN-Transformer-GAN model’s prediction performance across four building energy datasets: BIM-ECA, NEBULA, BEC, and BIM-BEM. The model consistently performs well, with NEBULA achieving the best results, recording the lowest RMSE (12.02 kWh) and highest R2 (0.934). The BIM-ECA dataset shows strong performance with RMSE of 14.32 kWh and R2 of 0.912, while the BIM-BEM dataset exhibits similar metrics (RMSE: 13.44 kWh, R2: 0.927). These results confirm the model’s effectiveness in accurately predicting energy consumption across diverse building configurations and datasets.
Tables 6 and 7 present the ablation experiments, and the results reveal the significant impact of the GNN and GAN modules on energy efficiency. From the baseline model to the final integrated model, both energy savings and energy utilization show consistent and substantial improvements. In the baseline model, the energy savings across four datasets range from 61.69% to 63.37%. After introducing the GNN module, the energy efficiency immediately increases to 75.71% to 77.31%, an average improvement of approximately 14 percentage points. When the GAN module is introduced alone, the energy savings further rise to 82.15% to 88.7%, with a more significant increase. When both GNN and GAN work together, energy savings reach an impressive 88.89% to 93.42%, and energy utilization also reaches 91.86% to 92.72%, nearly a 30 percentage point improvement compared to the baseline model. This consistency across datasets indicates that the GNN effectively captures complex network dependencies, while the GAN optimizes energy distribution strategies through adversarial mechanisms. The synergistic effect of both modules significantly enhances the performance of the building energy management system, providing an innovative solution for green buildings and energy conservation. Finally, the visualization of the results from Tables 6 and 7 is shown in Fig 9 below:
Tables 8 and 9 present the significant improvements in computational performance during the model architecture evolution. From the baseline model to the gradual introduction of the GNN and GAN modules, training time, inference time, and model parameters all show a consistent downward trend. In the BIM-ECA dataset, the baseline model’s training time is 53.41 seconds. After introducing the GNN, it decreases to 49.8 seconds, and with the addition of GAN, it further reduces to 42.18 seconds, representing a 21.02% reduction in training time compared to the baseline model. Inference time drops from 144.26 milliseconds to 110.23 milliseconds, and the number of parameters decreases from 2.6826 million to 2.2244 million. A similar trend is observed across other datasets. The synergy between the GNN and GAN modules not only improves the model’s energy management performance but also significantly optimizes computational efficiency. This performance boost stems from the GNN’s ability to capture network structural features more efficiently, while the GAN streamlines the model architecture through adversarial training mechanisms. The combined effect of both modules allows the model to maintain high accuracy while effectively utilizing computational resources, providing a more lightweight and intelligent technical solution for building energy management. The visualization of the results from Tables 8 and 9 is shown in Fig 10 below:
This study provides a comprehensive validation of the proposed GNN and GAN-based building energy management model through in-depth experimental analysis. The experimental results not only demonstrate the model’s significant advantages across multiple datasets but also reveal the unique value of the GNN and GAN modules in enhancing energy efficiency and optimizing computational performance. From energy savings to computational resource consumption, the proposed method achieves revolutionary breakthroughs compared to existing technologies, with energy savings increasing nearly 30 percentage points over the baseline model, and significant reductions in both training and inference times. More importantly, through the synergy of network graph learning and adversarial generation, the model successfully addresses the performance bottlenecks of traditional building energy management systems in complex scenarios, offering a more intelligent and efficient technical paradigm for smart buildings. The research enriches the theoretical foundation of innovative building energy management methods and demonstrates exceptional technical potential in practical applications, with significant implications for promoting green building development and improving urban energy utilization efficiency. Future work may further explore the model’s applicability to different building types and complex environments, continuously optimizing algorithm performance and contributing to the realization of low-carbon, efficient building energy management goals through ongoing technological innovation.
