Deakin University
Browse

File(s) under permanent embargo

Building Thermal Load Prediction Using Spatio-Temporal Graph Neural Networks

conference contribution
posted on 2023-02-26, 22:15 authored by Yilong JiaYilong Jia, Jun Wang, M Reza Hosseini, Wenchi Shou, Raphael Lin
Building thermal load prediction plays an essential role in reducing building energy consumption. However, most of the existing machine learning-based methods can only predict the thermal load of a single zone or the whole building, and it is difficult to perform multi-zone prediction. This study proposed a graph-based method to achieve simultaneous thermal load prediction for multiple zones. We firstly abstracted a graph data structure from the thermal zone layout and selected nine features for the graph nodes. Then, we designed a spatiotemporal graph neural network (ST-GNN) model consisting of graph convolutional network (GCN) modules and gated recurrent unit (GRU) modules. The GCN module captures the spatial relations of the graph nodes to learn the heat transfer between thermal zones, and the GRU module captures the temporal dependency of the embeddings to learn the change in thermal load over time. Finally, the proposed method was trained and tested using a simulated dataset of a small office building. The results show the proposed method is applicable with the average accuracy in the mean square error of 0.0829, the absolute error of 0.1761 and the coefficient of determination (R2) of 79.60%. The model performs well in predictions for eastern and western thermal zones. However, we found the accuracy for the sun-facing zone is relatively low and discussed the possible reasons. As a preliminary study, we also discussed the limitations of the proposed method and improvement solutions. This study demonstrates the feasibility of applying graph and graph neural networks in building energy and provides a novel machine learning model for multi-zone thermal load prediction.

History

Pagination

631-638

Location

Seoul, South Korea

Start date

2022-11-16

End date

2022-11-18

ISBN-13

978-0-9927161-4-1

Notes

Building thermal load prediction plays an essential role in reducing building energy consumption. However, most of the existing machine learning-based methods can only predict the thermal load of a single zone or the whole building, and it is difficult to perform multi-zone prediction. This study proposed a graph-based method to achieve simultaneous thermal load prediction for multiple zones. We firstly abstracted a graph data structure from the thermal zone layout and selected nine features for the graph nodes. Then, we designed a spatiotemporal graph neural network (ST-GNN) model consisting of graph convolutional network (GCN) modules and gated recurrent unit (GRU) modules. The GCN module captures the spatial relations of the graph nodes to learn the heat transfer between thermal zones, and the GRU module captures the temporal dependency of the embeddings to learn the change in thermal load over time. Finally, the proposed method was trained and tested using a simulated dataset of a small office building. The results show the proposed method is applicable with the average accuracy in the mean square error of 0.0829, the absolute error of 0.1761 and the coefficient of determination (R2) of 79.60%. The model performs well in predictions for eastern and western thermal zones. However, we found the accuracy for the sun-facing zone is relatively low and discussed the possible reasons. As a preliminary study, we also discussed the limitations of the proposed method and improvement solutions. This study demonstrates the feasibility of applying graph and graph neural networks in building energy and provides a novel machine learning model for multi-zone thermal load prediction.

Editor/Contributor(s)

Lee D

Title of proceedings

Proceedings of the 22nd International Conference on Construction Applications of Virtual Reality

Event

The 22nd International Conference on Construction Applications of Virtual Reality (CONVR2022)

Publisher

Chung-Ang University

Usage metrics

    Research Publications

    Categories

    No categories selected

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC