Spatio-Temporal Characterization and Short-term Prediction of Indoor Temperature in Multi-zone Buildings

Heating, ventilation, and air conditioning systems (HVAC) are essential for controlling indoor temperature and ensuring adequate levels of thermal comfort and indoor air quality. Among all environmental parameters, air temperature is certainly one of the most important. In-depth spatio-temporal characterization of this parameter has a significant impact on defining HVAC control strategies, allowing for optimization of building energy consumption while minimizing indoor comfort violations. In particular this study aims at developing a prediction model to forecast the indoor temperature values associated with 63 sensors installed in a real multizone office building located in California for 3 timesteps ahead, exploiting both temporal and spatial features extracted from long term monitored data. To this purpose, a Spatial–Temporal Graph Neural Network (STGNN) algorithm is used, and its performance is compared against a baseline Long Short-Term Memory (LSTM) neural network model. The results demonstrated that although the performance of the two approaches is generally comparable, the GNN-based model is able to achieve a more accurate result and to better estimate temperature values at longer time horizons. In addition, the STGNN model allowed the development of a single multi-output prediction model, generating a significant advantage from the implementation point of view.