This work proposes a generalizable thermal dynamics modeling method by coordinating multiple buildings. It formulates a federated learning framework to facilitate collaborative modeling in a privacy-preserving way. It further identifies dual heterogeneity in both model structures and data distributions for federated learning, and then proposes a two-level personalization strategy combining similarity matrices and adaptive weighting to alleviate the impacts.
Codes for submitted Paper "Generalizable Thermal Dynamics Modeling via Personalized Federated Learning".
Authors: Xueyuan Cui, Dalin Qin, Jean-François Toubeau, François Vallée, and Yi Wang.
Python version: 3.8.17
The must-have packages can be installed by running
pip install requirements.txt
All the data inputs for experiments can be acquired in Data here and Data_in that can be downloaded from Google Drive.
To reproduce the experiments of the proposed methods and comparisons in the paper, please go to the folder
cd #Codes
where the introduction on the running order and each file's function is explained in readme.md.
Note: There is NO multi-GPU/parallelling training in our codes.
All models that are trained by the proposed method and other comparisons can be acquired in the folder #Results that can also be downloaded from Google Drive.
In particular, Models_newcomb, Models_oneW, Models_fedavg, and Models_single includes models from "DW_ada", ("DW_fix"&"SW_data"&"SW_model"), "SW_avg", and "Local" methods, respectively.
Package #Codes/torchdiffeq1/ is modified based on the open code of Neural ODE. The rapid development of this work would not have been possible without this open-source package.