Paper Conference

Proceedings of Building Simulation 2023: 18th Conference of IBPSA

     

Graph Convolutional Network (GCN) for predicting multizone airflow path

Xiaoshi Wang 1,2, Xu Han 1,2, Ali Malkawi 1,2, Na Li 3
1 Harvard University Green Center for Building and Cities
2 Harvard Graduate School of Design
3 Harvard School of Engineering and Applied Sciences (SEAS)


DOI: https://doi.org/10.26868/25222708.2023.1413
Abstract: Indoor airflow path is important for multizone indoor space ventilation condition and contaminant control. Traditionally, Computational Fluid Dynamics and Airflow Network model are mainly used to calculate airflow path. CFD is challenged by its high computational cost and Airflow Network model can not reflect the influence of flow pattern distribution. This paper presents a novel machine learning model to predict multizone indoor airflow path with comparable accuracy of CFD generated result in a fast-responding way. The model uses Graph Convolutional Network, which makes prediction based on information stored in arbitrarily structured graphs. The model has 91% accuracy in predicting airflow direction for randomly generated 5-room indoor space.
Keywords: Graph Convolutional Network, multizone airflow path, CFD result dataset
Pages: 1031 - 1037
Paper:
bs2023_1413