Electricity data sensors are widely used across large buildings and households. As the data is collected by distributed sensors from varied locations, privacy-preserving becomes a top concern for data owners. Meanwhile, multiple deep learning models achieved state-of-art performance on forecasting with the electricity time series data in a centralized training mechanism. Although these deep learning models are powerful at capturing temporal features and making precise predictions, it usually consumes a large amount of memory and resources during the training process. To address two problems, i.e., the data privacy issue and high-demanded resources for training, we propose an efficient and practical deep learning model using a transformer framework while utilizing federated learning to move the training on local data instead of on a centralized place. With the proposed deep learning model, the computation will reduce its memory usage by 60% while achieving similar and even better results on forecasting with the electricity time series data. Case studies on the university communities’ building demonstrate our proposed solution’s great potential and comparative performance compared to the state of the arts.
|