A framework for demand response load forecasting based on federated learning

Abstract Addressing the scenario where distribution network operators predict load curves based on user data to participate in demand response, and to cope with issues such as user data privacy protection and the operators’ lack of permission to access metering data, this paper proposes a framework and method for demand response load forecasting based on federated learning. Focusing on features including weather, time, and users’ participation status in demand response, this approach utilizes horizontal federated learning technology to integrate data from various sources while safeguarding user data privacy. Furthermore, a balanced scheduling scheme for participating users and samples, along with data preprocessing, is integrated to enable effective discrimination of load curves. Case study analysis demonstrates that the proposed load forecasting method can maintain superior prediction performance while ensuring the privacy of user data.