Transformer-based energy consumption prediction and optimization framework for BIM-enabled green buildings

Traditional building energy management systems rely on physics-based simulations or data-driven forecasting methods that fail to leverage rich semantic information encoded in Building Information Models (BIM), resulting in substantial performance gaps between design intent and actual operations. This paper proposes an intelligent energy prediction and optimization framework based on Transformer architecture with BIM-enabled multi-modal fusion. The system integrates five heterogeneous data sources—building operations, indoor-outdoor environment, weather-season, BIM deviceoccupancy, and BIM geometry-semantics—into a unified 256-dimensional embedding space through modal-specific projections. A novel BIM-guided cross-modal attention mechanism enables static architectural characteristics to modulate interpretation of dynamic sensor streams, capturing building-specific energy response patterns. The model employs multi-task learning to simultaneously predict energy efficiency ratings (5-level classification), total consumption, peak load timing, and subsystem breakdowns. Experimental results on 18,500 samples from three operational green buildings demonstrate that the proposed method achieves 8.23% MAPE and 0.921 Macro-F1 score, representing 23.5% error reduction compared to Attention-LSTM baseline and 43.5% improvement over XGBoost. The hybrid edge-cloud deployment architecture achieves real-time inference latency below 2 seconds through model compression while maintaining prediction accuracy within 5% of the full model. This work validates the effectiveness of incorporating BIM domain knowledge into deep learning frameworks for enhanced building energy management.