Comparison of different deep neural networks for system identification of thermal building behavior

Having accurate information available about future thermal building behavior can help to make good decisions in various heating control tasks. However, creating precise mathematical models for many different buildings is a complex and time-consuming task, owing to the heterogeneity of the building stock and the behavior of its occupants. In this paper, we propose a DNN-based system identification approach for predicting the room temperature inside a building based on past information and future weather forecasts. We evaluate various state-of-the-art and custom-built DNN architectures for TSF. Besides prediction performance, storage space and inference speed as measures for the respective model's complexity are also taken into account. Our main contribution is demonstrating the effectiveness of these models in predicting the room temperature for differently parameterized simulated buildings. By using several distinct buildings for training, validation and testing, we additionally show that these models are capable to generalize in a way such that the room temperature for different buildings can be predicted by a single model, without any changes or adaptions.