Convolutional neural network to learn building-shape representations for early-stage energy design

Abstract

Machine learning (ML) supports energy-efficient design by quickly predicting the building performance at early stages. However, existing ML approaches cannot easily learn geometrical representations and their interactions with technical specifications for predicting energy. Common approaches of artificial neural networks (ANN) use either numerical parameters to represent building geometry or component-based ML (CBML) to train transferrable ML components. However, both approaches have high generalisation errors for complex geometries. We used a convolutional neural network (CNN) to learn geometrical representations of buildings and an autoencoder to optimise the training process. We test existing approaches of ANN and CBML with a proposed approach of CNN for different sizes of training datasets. For the biggest training dataset, ANN, CBML, and CNN have generalisation errors (root-mean-squared errors) of 1.35, 1.60, and 0.77 kWh/a.m2, respectively. While CBML has lower prediction errors on small datasets, CNN is more suitable for large datasets. Also, CNN most often has better prediction accuracy than ANN and trains faster on encoded data. Since CNN captures the effects of building geometry on energy predictions better than ANN and CBML approaches, it is an attractive choice for early-stage energy design.