Neural Networks as Metamodel for Hygrothermal Simulations of Building Components – Reducing the Calculation Time of Probabilistic Assessments

Simulating the hygrothermal response of a building component often involves many uncertainties, such as the exterior and interior climate, or even the exact geometry and material properties. A deterministic assessment often does not suffice to come to a reliable design decision or conclusion, whereas a probabilistic evaluation includes these uncertainties, and thus allows assessing the hygrothermal behaviour and the related damage risks more reliably. However, the thereto frequently adopted Monte Carlo approach often involves thousands of simulations, and therefore easily becomes computationally inhibitive. To overcome this issue, the hygrothermal model can be replaced by a metamodel, a simpler and faster mathematical model, mimicking the original hygrothermal model, thus strongly reducing the calculation time. Static metamodels have already been applied in the field of building physics multiple times. The main disadvantage is that these types of metamodels are developed for a specific single-valued performance indicator (e.g. the total heat loss or the maximum mould growth index). If the user wishes to evaluate the performance using a different indicator a new metamodel needs to be constructed, which is time intensive. Additionally, single-valued performance indicators provide less information, which might impede decision-making. Instead, a dynamic metamodel predicting the hygrothermal output time series, as calculated by the original hygrothermal model (temperature, relative humidity, moisture content …), supports a more flexible approach. However, the pattern between input time series (e.g. exterior climate, interior climate) and output time series (e.g. temperature, relative humidity, moisture content …) are much more complex to model, due to its highly non-linear and transient nature. The metamodel must thus be able to capture these complex and time dependent patterns and therefore, not all metamodelling strategies are suited for time series prediction. Hence, the overall aims of this research are (1) the development a time-efficient metamodel that can accurately predict the hygrothermal time series and (2) its illustrative application on the hygrothermal evaluation of a building component.

To develop an accurate and time-efficient metamodel, this research focuses on artificial neural networks, as they have been applied successfully for modelling complex and non-linear sequences in many other fields. A comparison of several neural network types showed that both recurrent neural networks (Long-Short Term Memory and Gated Recurrent Unit) and convolutional neural networks are able to accurately predict the hygrothermal response of a massive brick wall based on the exterior and interior climate time series. Of these two network types, the convolutional neural network is found to be significantly faster in both training and predicting, and thus is deemed most suited as metamodel. Hence, the convolutional neural network is further developed and fine-tuned to predict the hygrothermal response of a massive masonry wall. Furthermore, a strategy to optimise the network’s hyper-parameters efficiently is developed, based on the Grey-Wolf algorithm.

At the end of the thesis, the use of a convolutional neural network as metamodel is illustrated for the hygrothermal evaluation of timber frame walls. The developed convolutional neural network is adapted for the intended modelling application and the networks’ hyper-parameters are optimised using the proposed strategy. This results in a metamodel that is 500 times faster than the original hygrothermal model. Finally, the network is used to calculate the hygrothermal response of 96 different timber frame walls, taking into account all influencing uncertainties. This data subsequently allows a reliable comparison between the different wall compositions and allows formulating recommendations.