A hybrid approach to thermal building modelling using a combination of Gaussian processes and grey-box models

This paper presents a hybrid building modelling method with a reduced modelling and calibration effort. The method combines a physics-based model, which describes the general behaviour of the system, with a machine learning algorithm trained to correct the physics-based model's systematic errors. To exemplify the method, a highly simplified grey-box model is used as the physics-based part and a Gaussian process as the machine learning part. It is shown that the hybrid model improves the temperature and energy predictions of the grey-box model while having a lower generalization error than the pure Gaussian process. Specifically, the hybrid approach achieved a day-ahead zone temperature prediction error ca. 0.1 K (RMSE) lower than the grey-box model. As for the energy prediction, the hybrid model obtained an error of 3% compared to 8% for the grey-box model. In comparison to the Gaussian process, the hybrid approach achieved better predictions in all cases. The improvements were especially high when the models were trained with small datasets: 0.7 K in the temperature prediction and 25 percentage points in the energy prediction.