Personalized hierarchical heterogeneous federated learning for thermal comfort prediction in smart buildings

Federated Learning (FL) is gaining significant traction due to its ability to provide security and privacy. In the FL paradigm, the global model is learned at the cloud through the consolidation of local model parameters instead of collecting local training data at the central node. This approach mitigates privacy leakage caused by the collection of sensitive information. However, it poses challenges to the convergence of the global model due to system and statistical heterogeneity. In this study, we propose a two-fold Personalized Hierarchical Heterogeneous FL (PHHFL) approach. It leverages a hierarchical structure to handle statistical heterogeneity and a normal distribution-based client selection to control model divergence in FL environment. PHHFL aims to use a maximum number of local features of each client and assign specific level in the hierarchy. Furthermore, to address model divergence caused by the nodes’ statistical heterogeneity, we propose a novel client selection strategy based on the performance distribution of the nodes. Experiments are conducted on thermal comfort datasets and a synthetic dataset with 12 and 10 clients, respectively. The results show that the proposed PHHFL outperforms in terms of accuracy, F1 score, and class-wise precision on both thermal comfort and synthetic datasets. The source code of the PHHFL model and datasets is available on GitHub.