Design information-assisted graph neural network for modeling central air conditioning systems

Buildings consume huge amounts of energy to create a comfortable and healthy built environment for people. The building engineering industry has benefitted from the advances in building informatics, including the rich data available in modern buildings and the rapid development in computing technology and data science, for building energy management. Dynamic modeling is often essential to online control and optimization of building energy systems. Data-driven modeling empowered by advanced machine learning has achieved ground-breaking performance in capturing temporal relationships among multivariate building operation data in recent years. However, the structural relationships among the physical entities, e.g., the topology of air conditioning ductworks and terminals, are generally overlooked in existing data-driven modeling methods, although they are very helpful in capturing the relationships among building operation data. This study proposes to represent building air conditioning systems as graphs for machine learning, whose nodes and edges represent physical entities (e.g., VAV terminals) and their connections (e.g., ductwork), respectively. A novel graph neural network-based methodology is developed for dynamic modeling of central air conditioning systems, which consists of three steps, i.e., automated graph structure design, development of graph neural network, and model evaluation and explanation. Viable and generalizable graph structure design methods based on design information, e.g., design drawings and BIM models, and machine learning algorithms for model development and explanation are proposed. A case study of dynamic modeling of a real central air conditioning system serving the tallest building in Hong Kong is carried out by adopting the methodology developed. Image identification techniques are employed to extract the topology of the air conditioning system from 2D schematic drawings, which is used as the prototype of the graphs. Graph neural network-based models, consisting of a graph layer and a recurrent layer, are developed to capture the structural and temporal relationships, separately. The graph layer learns structural relationships from the input graphs by using graph convolutional network (GCN) and graph attention network (GAT). The recurrent layer learns temporal relationships from massive historical operation data by using LSTM. As the models become much "darker", the Shapley additive explanations (SHAP) method is adopted to provide both global and local model explanations, i.e., the impact of each model input on outputs. The developed models exhibit improved capabilities of automated model architecture design, prediction accuracy, generalizability and interpretability. The methodology developed for dynamic modeling of air conditioning systems in this study leverages both traditional building system design information and powerful machine learning algorithms, and exemplifies an ideal synergy between engineering domain expertise and machine learning.