Personalized federated learning for household electricity load prediction with imbalanced historical data

Household consumption accounts for about one-third of global electricity. Accurate results of household load prediction would help in energy management at both the building and the grid levels. Data-driven household load prediction methods have shown great advantages and potential in terms of accuracy. However, these methods still face challenges such as limited data for individual households, diversified electricity consumption behaviors, and data privacy concerns. To solve these problems, this paper proposes a personalized federated learning household load prediction framework (PF-HoLo), which allows personal models to learn collectively, leverages multisource data to capture diverse consumption behaviors, and ensures data privacy. In addition, the global encoder model and mutual learning are proposed to enhance the performance of the PF-HoLo framework considering imbalanced residential historical data. Ablation experiments results prove that the PF-HoLo framework could achieve significant improvements, with 13.41% Mean Square Error and 11.33% Mean Absolute Error, compared to traditional federated learning methods. • PF-HoLo predicts household loads while retaining features unique to imbalanced data. • Only Encoder parameters are shared to maximize shared knowledge and preserve privacy. • Hidden state soft targets enhance the capability of personalized federated learning. • Experiments validate the effectiveness under imbalanced data with ablation studies.