A scalable and transferable approach for spatio-temporal indoor air temperature forecasting in multi-zone buildings exploiting knowledge graph contextual information

This paper presents a scalable and transferable approach for spatio-temporal indoor air temperature forecasting in multi-zone buildings, utilizing contextual information from a knowledge graph. The method extracts static and dynamic embeddings for devices and zones from the knowledge graph and timeseries data, enabling flexibility across diverse building configurations. By leveraging these embeddings, the model can effectively handle varying HVAC setups and predict temperature evolution while exploiting only half of the training dataset. The approach achieves a mean absolute error (MAE) of 0.4 for a 24-hour prediction horizon over a 6-month test period, demonstrating comparable performance to state-of-the-art methods. Its main advantage lies in its scalability and transferability, as the use of knowledge graph embeddings allows the model to capture contextual information while learning shared thermal patterns.