Power Data Analysis and Privacy Protection Based on Federated Learning

Abstract

In order for active distribution network operators to carry out power business such as load forecasting without meter reading rights, the author proposes a research on power data analysis and privacy protection based on federated learning. The author proposes a federated learning load forecasting framework for industry user data protection by selecting weather and time factors as the correlation factors of load. On this basis, the author constructed an industry user dataset and established a load forecasting model based on Long Short Term Time Series Network (LSTNet). At the same time, the FedML framework was used to establish a sub industry load forecasting framework based on federated learning. The results indicate that: The accuracy of the industry specific load forecasting framework based on federated learning proposed by the author is less than 9 p.u., and the theoretical maximum value of SMAPE is 210 p.u., indicating that this method has universality and universality and can be applied in different industries. The training method of this scheme is parallel, although it increases the interaction time by 1 minute, the interaction time accounts for a smaller proportion compared to the training time, and the time consumption is interaction time (1 minute)+single training time (94 minutes). Conclusion: The method can enable users in the same industry to conduct federated training without sharing load data, and support active distribution network operators in related business operations while protecting user electricity privacy. It has better predictive performance, fewer model numbers, and shorter time consumption.