Graph convolutional networks-based method for estimating design loads of complex buildings in the preliminary design stage

Data-driven methods have shown powerful capabilities in estimating design cooling, heating and electricity loads of general buildings during the preliminary design stage. However, it is still quite challenging for them to handle complex buildings. Since conventional methods cannot describe the complex structures properly in a mathematical way. To address this challenge, this study proposes a graph convolutional networks-based method. The method includes three steps, i.e., graph representation, graphical model learning, and graphical model interpretation. Graph representation is designed to divide complex buildings into several basic blocks and utilizes graphs to represent the relationships among blocks. A graph convolutional network model, a state-of-art graphical model, is then trained to learn the energy load patterns of each basic block using the graphs. After that, class activation mapping, an effective interpretation algorithm for graph convolutional networks, is applied to quantify the contribution of each building feature to the model output. Operational data of 800 simulated complex buildings with significantly different structures are generated using SketchUp and Openstudio to verify the performance of the proposed method. Four representative data-driven models (artificial neural networks, random forest, support vector regression and gradient boosting tree) and one fast physical principle-based estimation method are selected as baseline methods for performance comparison with the proposed method. Results show that the proposed method has the highest accuracy. According to the results of class activation mapping, it is also demonstrated that the knowledge learned using a graph convolutional network is consistent with the domain knowledge.

Automated summary

A graph convolutional networks-based method is proposed to estimate the energy loads of complex buildings, showing the highest accuracy among various data-driven models. The method divides buildings into basic blocks, learns energy load patterns using graphs, and interprets the relationships between building features and model output using class activation mapping.