Privacy-Preserving Power Consumption Prediction Based on Federated Learning with Cross-Entity Data

Big data analytics have become vastly popular in the power system. However, power-related datasets are often owned by different parties, and a centralized data mining scheme in the form of direct data sharing might compromise party benefits, individual privacy and even national security. To this end, we implement a systematic privacy-preserving federated learning framework in the power system, which enables collaborative learning of power consumption patterns with cross-entity data. Horizontal federated learning is employed when data are scattered by samples; vertical federated learning, on the other hand, is designed for the case with data scattered by features. Further, we propose an alternative encryption scheme based on the Diffie-Hellman key exchange protocol to greatly boost computational efficiency. Case studies show that models trained from the proposed federated learning framework are lossless, privacy-preserving, efficient and effective.