Neural networks to predict the hygrothermal response of building components in a probabilistic framework

In recent years, probabilistic assessment of hygrothermal performance of building components has received increasing attention. Given the many uncertainties involved in the hygrothermal behaviour of building components, a probabilistic assessment enables to assess the damage risk more reliably. However, this typically involves thousands of simulations, which easily becomes computationally inhibitive. To overcome this time-efficiency issue, this paper proposes the use of much faster metamodels. This paper focusses on neural networks, as they have proven to be successful in other non-linear and non-stationary research applications. Two types of networks are considered: the traditional multilayer perceptron (with and without a time window) and memory neural networks (LSTM, GRU). Both are used for predicting the hygrothermal behaviour of a massive wall. The results showed that all networks are capable to predict the temperature profiles accurately, but only the LSTM and GRU networks could predict the slow responses of relative humidity and moisture content. Furthermore, the LSTM and GRU network were found to have almost equal predicting accuracy, though the GRU converged faste