Supply Air Temperature Control in Air Handling Unit based on Federated Learning

The study presents the application of a graph-based Federated Learning (FL) approach to model the supply air temperature in an air handling unit (AHU). An empirical graph including four machines was constructed based on Kullback-Leibler Divergence metrics, and the high-weight edges were used for model training by an FL algorithm: FedSGD. Each graph node includes a local data set consisting of the ambient temperature and the average of the room temperatures at which each AHU is operating. The study highlights the results of the application of FedSGD in cases where all local dataset features are present and when some of them are missing. Experimental results demonstrate a slight outperformance of the local linear regression models for all present features compared to the FedSGD method. A positive tendency in the application of FedSGD can be noticed for the nodes with missing features. While studying the impact of the missing features based on the learning performance, it appears that the absence of ambient temperature has a higher effect compared to the average room temperature.