Household profile identification for retailers based on personalized federated learning

With the deployment of smart meters, the retailer could obtain household profile information from massive data and implement demand response. However, different retailers in the retail market could not share the consumer's electricity consumption data as these data are regarded as their own assets and cause the problem of data island. In the meanwhile, the electricity data are highly heterogeneous since different retailers may serve various consumers and retailers prefer to acquire personalized models. To this end, a personalized household profile identification framework based on federated learning is proposed, which could protect the retailer's privacy and satisfy the individual requirement as well. Specifically, hypernetwork with an embedding layer is implemented on the server's side as a personalized technique, where individual models are generated during the training process. Case studies show that the proposed method has similar performance to the conventional centralized training method even if privacy protection is considered and outperforms the method applying a traditional federated learning averaging algorithm. Besides, an experiment with a new retailer participation is designed and conducted to verify the effectiveness of the unique design in our method. © 2023 Published by Elsevier Ltd.