The construction field is known for its high professional requirements, complexity, and diversity. Over the past few decades, CO₂ emissions and energy consumption have increased significantly, making it necessary to improve the efficiency of building design and the retrofitting for energy conservation imperatively. Pre-trained large language models (LLMs) assist industry professionals in architectural design and renovation through natural language processing. This study reviews research papers on Transformer architecture pre-training LLMs from the perspective of model types and application directions. It discusses their potential future application directions along with factors affecting their performance. The assessment outcomes indicate that LLMs excel in tasks including building information detection, modeling automation, and fault diagnosis. Specifically, LLMs achieve median accuracy of approximately 89% for building information detection, 91% for auxiliary modeling, and 98% for fault detection. However, current applications are primarily limited to processing built environment data (e.g., structural data and operational/maintenance records). The input prompt is an essential factor that affects the task execution capability of LLMs. Existing studies have not fully explored the interpretability, transfer learning potential, or contextual learning methods of LLMs, and most research is limited to the use of GPT models. Future work should expand the applications of LLMs in construction, develop task-specific model adaptations, enhance model interpretability, and establish new evaluation metrics for construction-associated tasks.

The prediction of building energy consumption offers essential technical support for intelligent operation and maintenance of buildings, promoting energy conservation and low-carbon control. This paper focused on the energy consumption of heating, ventilation and air conditioning (HVAC) systems operating under various modes across different seasons. We constructed multi-attribute and high-dimensional clustering vectors that encompass indoor and outdoor environmental parameters, along with historical energy consumption data. To enhance the K-means algorithm, we employed statistical feature extraction and dimensional normalization (SFEDN) to facilitate data clustering and deconstruction. This method, combined with the gated recurrent unit (GRU) prediction model employing adaptive training based on the Particle Swarm Optimization algorithm, was evaluated for robustness and stability through k-fold cross-validation. Within the clustering-based modeling framework, optimal submodels were configured based on the statistical features of historical 24-hour data to achieve dynamic prediction using multiple models. The dynamic prediction models with SFEDN cluster showed a 11.9% reduction in root mean square error (RMSE) compared to static prediction, achieving a coefficient of determination (R²) of 0.890 and a mean absolute percentage error (MAPE) reduction of 19.9%. When compared to dynamic prediction based on single-attribute of HVAC systems energy consumption clustering modeling, RMSE decreased by 12.6%, R² increased by 4.0%, and MAPE decreased by 26.3%. The dynamic prediction performance demonstrated that the SFEDN clustering method surpasses conventional clustering method, and multi-attribute clustering modeling outperforms single-attribute modeling.