Estimation of parameter values in heat pump systems using neural networks
Heat pump systems are a key component in the decarbonization of the building sector. However, efficient control of these systems requires technical parameters that are often unknown or undocumented in real-world installations. Relevant parameters include pipe length, pipe inner diameter, and buffer storage volume. This paper investigates the use of artificial neural networks (ANNs) to estimate these parameters based on timeseries data from selected sensor signals. Several network architectures are examined, including Multi-Layer Perceptrons (MLPs), Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), and Generative Adversarial Networks (GANs). The models are trained on a simulation model of a building energy system comprising a heat pump, a buffer storage tank, a domestic hot water tank, and the relevant hydraulic components. By varying the pipe length, pipe inner diameter, and buffer storage volume, different system configurations are simulated. The resulting sensor time series serve as input to the networks, while the corresponding system parameters represent the target outputs. In the first study, the models’ ability to estimate individual parameters is investigated. The CNN and RNN achieve the lowest mean errors, whereas the GAN exhibits the highest error values. In the second study, the simultaneous estimation of all three parameters is analyzed. Again, the CNN and RNN achieve the highest accuracy, with mean errors of 0.8 % for pipe length, 1.7 % for pipe inner diameter, and 0.9 % for buffer storage volume. The results highlight the potential of neural networks for data-driven identification of technical parameters in building energy systems. Future research should investigate the transferability of the proposed approach to real measurement data and explore methods for uncertainty quantification.