Moisture prediction of timber for durability applications using data-driven modelling

Durability and service life assessment is a major challenge for the design and use of timber in outdoor weather exposed environments. Rate of deterioration by fungal decay is closely linked to variations in wood moisture content. The objective of the present paper is to test and evaluate different data-driven models based on the multilinear regression (MLR) and artificial neural network (ANN) approach. Moisture content was predicted at the surface and core of a rain-exposed wooden element in the context of durability and service life assessment. Synthetic data stemming from a numerical model were used to fit time-series weather variables, including different combinations of time-lagged daily precipitation, relative humidity, and temperature, to temporal variations of daily average wood moisture content. Based on a set of statistical and qualitative analyses, using the weather variables lagged by 0 – 11 days as input variables for 11 mm depth moisture prediction, ANN showed the highest accuracy and least sensitivity to its initial setups, and could significantly outperform the MLR with the same input variables. The resulting models for surface and core moisture prediction were then tested against two different datasets consisting of measured data from wood specimens subjected to outdoor exposure.