Online federated learning based short-term load prediction through anomaly detection using incremental PCA

Abstract

In this paper, we propose an incremental principle component analysis (IPCA)-based federated learning (FL) framework for day-ahead load forecasting, where IPCA is utilized to reflect seasonal variation and detect anomalies when training models in local devices. FL is a promising learning scheme for leveraging distributed data and computational resources while preserving user privacy. However, FL is susceptible to anomalies in local devices, which can degrade the overall performance of the global model. These anomalies can include corrupted data, adversarial attacks, or simply unreliable data. In particular, in the case of power consumption profiles, seasonal variations can mask such anomalies because the deviation of anomalies can be smaller than that of seasonal variations. To address this issue, in our framework, mini batches for training models in local devices are transformed by IPCA matrices into low dimensional spaces, where Mahalanobis distances between previous and current mini batch are calculated to determine whether re-initialization is needed by seasonal variation or or anomaly is declared for the current mini batch. To further improve prediction performance, we customized the IPCA-FL by adding an online learning scheme with greedy – i –selection algorithm for model updates. We evaluated the two proposed schemes (offline IPCA-FL and online IPCA-FL with greedy selection) on Gwangju Institute of Science and Technology (GIST) building electricity load dataset (i.e., research bldg. 8). The results showed that our schemes outperformed traditional centralized learning and existing FL methods in terms of root mean square error (RMSE) and mean absolute percentage error (MAPE) metrics. Moreover, under various scenarios with less than one year of data, federated learning showed superior performance compared to individual learning. however, it was observed that individual learning outperformed our methods when the training data is long enough to include seasonal variations.