A gap-filling method for room temperature data based on autoencoder neural networks

This study explores the applicability of a deep learning-based approach for reconstructing missing room temperature data from different domains where relatively few training samples are available. For that purpose, the existing convolutional, long short-term memory (LSTM) and feed-forward autoencoders were combined with a suitable domain adaptation procedure. Eventually, the developed models were evaluated on data collected in four buildings with significant differences in thermal mass, design and location. The findings pointed out that the domain adaptation can be conducted effciently by using a small data sample from the target domain. Additionally, the results showed that the proposed model can reconstruct up to 80 % of the missing daily room temperature inputs with RMSE accuracy of 0.6 °C.