Analyzing Parameter Estimation Methods for RC Models in the Modelling of Heat Dynamics of Residential Buildings

The growing use of heat pumps contributes significantly to energy consumption and is thus a suitable focus for flexibility solutions. Maximizing the flexibility of a single heat pump requires an accurate model of a building’s heat dynamics, ensuring that the heat pumps can be controlled appropriately. These dynamics are unique to each building and continuously evolve as a result of a variety of factors such as weather, home remodelling, and building usage. Resistor-capacitor (RC) models are a proven way to model such dynamics, but require parameters to be tuned to the building represented by the model. Training algorithms can be applied to learn these parameters based on observational data. This paper qualitatively compares three such algorithms: a genetic algorithm, a neural network and a sequential Monte Carlo algorithm. The quality of the learning algorithms is assessed on the accuracy, its adaptability, explainability, and the amount of time and data required to converge. This serves as the foundation regarding the practical usability of such an algorithm in a real-world environment.