Thermal building modelling using Gaussian processes

• A Gaussian process (GP) model is trained to predict the zone temperature in a building.
• The prediction error of the GP is compared to a physics-based grey-box model.
• The GP presented a high error when faced with data values out of the training range.
• Adaptive training is applied to the GP, reducing the prediction error in up to 50%.

This paper analyzes the suitability of Gaussian processes for thermal building modelling by comparing the day-ahead prediction error of the internal air temperature with a grey-box model. The reference building is a single-zone office with a hydronic heating system, modelled in TRNSYS and simulated during the winter and spring periods. Using the output data of the reference building, the parameters of a Gaussian process and of a physics-based grey-box model are identified, with training periods ranging from three days to six weeks. After three weeks of training, the Gaussian processes achieve 27% lower prediction errors during occupied times compared to the grey-box model. During unoccupied times, however, the Gaussian processes perform consistently worse than the grey-box model. This is due to their large generalization error, especially when faced with untrained ambient temperature values. To reduce the impact of changing weather conditions, adaptive training is applied to the Gaussian processes. When re-training the models every 24 h, the prediction error is reduced over 21% during unoccupied times and over 10% during occupied times compared to the non-adaptive training case. These results show that the proposed Gaussian process model can correctly describe a building's thermal dynamics. However, in its current form the model is limited to applications where the prediction during occupied times is more relevant.