Reinforcement Learning-Enhanced BIM-Based Physics-Constrained Graph Neural Network Energy Management Framework

Building operational energy consumption is one of the main sources of global energy consumption and carbon emissions. In the context of high-quality development and the ``dual carbon'' strategic goals, efficient and interpretable building energy consumption management methods hold significant theoretical and practical importance. Existing studies primarily rely on BIM energy consumption assessment based on physical simulation or single data-driven models, which suffer from shortcomings such as computational complexity, inability to capture complex nonlinear features, difficulty in considering dynamic feedback, and lack of interpretability. To address the aforementioned issues, this paper proposes a reinforcement learning-enhanced BIM-based physics-constrained graph neural network energy management framework. This framework extracts geometric, material, and topological features from the building information model, employs a physics-constrained graph neural network for energy consumption prediction, and integrates reinforcement learning to achieve real-time control strategies. Specifically, we fuse BIM features through tensor representation and graph convolution, design a physics-constrained loss function embedding the heat conduction equation, and construct a reinforcement learning policy network based on value iteration to achieve integrated energy consumption prediction and control. Extensive experiments on multiple public and self-built datasets demonstrate that this method significantly outperforms XGBoost, Random Forest, LightGBM, CatBoost, TabNet, AutoGluon, FT-Transformer, as well as recently proposed Physics-Informed GNN, BuildingGym reinforcement learning framework, Neural Operator Transformer, and other methods in terms of MAPE, RMSE, and other metrics, with an average error reduction of over 20\\%, and exhibits excellent generalization and robustness. The research results indicate that the deep coupling of reinforcement learning and physics-constrained graph neural networks provides new theoretical tools and practical pathways for building energy consumption management, promoting the application of BIM--AI integration in the energy field.