Tuning Performance of Grey-Box Models for Thermal Building Applications

The electrification of heating and cooling systems transforms energy consumption patterns, creating challenges and opportunities for power system operation. To this end, RC (or grey-box) models have emerged as promising modeling structures for forecasting the thermal needs of buildings. This paper presents a tuning algorithm and performance evaluation of the grey-box model using realistic measurement data. It provides insights into the sensitivity of the learning process to sub-optimal solutions and the computational burden. Four different grey-box structures with varying complexity are evaluated. The proposed methodology is critical in integrating heating and cooling systems into future power systems. The results reveal that the 4R3C model achieves the highest accuracy among the evaluated structures; however, it requires significantly higher computational time in certain scenarios. Moreover, the physical representation of the building through this training structure can be challenging, as the optimization process does not consistently converge to a unique set of parameter values, indicating the presence of multiple local optima. The tuning framework is provided as an open-source modeling tool, aiming to support further research on grey-box model.