Paper Conference

Proceedings of Building Simulation 2021: 17th Conference of IBPSA


CNN-based quick energy prediction model using image analysis for shape information

Manav Mahan Singh 1, Philipp Geyer 1,2
1 KU Leuven
2 TU Berlin

Abstract: Data-driven approaches are useful to substitute computationally expensive tools used for energy prediction. These approaches allow developing quick energy prediction models, essential to promote energy analysis at the early stages. However, developing a wellgeneralising model is challenging due to varying building shape. This article develops such a model using a deep learning approach. A convolutional neural network (CNN) with modified architecture is used to capture the shape and technical specifications. The model predicts energy use intensity with a mean-absolute-percentageerror of 1.51% and root-mean-square-error of 1.06 kWh/m2a. Integrated with building information modelling (BIM), the model predicts probabilistic energy performance for 5000 samples in 65 seconds to support informed decision-making under uncertain scenario.
Keywords: Deep Learning, Early Design Stage, Energy Simulation, Convolutional Neural Network, Image Analysis
Pages: 1311 - 1316