The impact of a reduced training subspace on the prediction accuracy of neural networks for hygrothermal predictions

Abstract

Performing a probabilistic assessment of a building component can easily become computationally inhibitive. To solve this issue, the hygrothermal model can be replaced by a metamodel, which mimics the original model with a strongly reduced calculation time. In this paper, convolutional neural networks are used to predict hygrothermal performance. Because neural networks do not extrapolate well outside their training subspace, it is important to select the training data wisely so that the network can be used to predict for a wide variety of cases, while keeping training time as low as possible. The impact of a reduced training subspace is investigated by training a network on a limited number of wall types or exterior climates and evaluate its prediction accuracy for different wall geometries or other climates. The results showed that is indeed possible to train on a well-considered reduced subspace, while maintaining high accuracy, though it does not necessarily save training time.