Robust Co-Modeling for Privacy-Preserving Short-Term Load Forecasting With Incongruent Load Data Distributions

Short-term load forecasting (STLF) can support strategies for power retailers. Training based on aggregate load datasets can yield more precise STLF models. In real scenarios, however, retailers can only access the consumer information they serve. The aggregation of their data for centralized forecasting requires access to private local data, which may lead to potential security concerns. This paper proposes a method namely Privacy-Preserving {k} -means Federated Learning (PPK-Fed), to facilitate STLF co-modeling among retailers. Federated Learning (FL) is founded on the assumption of data consistency, which does not necessarily exist in real datasets. PPK-Fed incorporates {k} -means clustering together with convolutional neural networks into FL. PPK-Fed is proven to reduce the impact of potential data incongruence embedded in retailer local datasets. Besides, to address the sensitivity of the {k} -means prototype to the potential anomalies in the real dataset, PPK-Fed fuses density-based anomaly detection into {k} -means clustering under FL to improve robustness. For further model security, a secure multi-party computation (SMPC) scheme is designed in PPK-Fed. The model validity, privacy-preserving features, and robustness to anomalies have been verified using a real load dataset.