Abstract: To effectively address the issues of secure data sharing and collaborative modeling between different power companies in the process of electricity load forecasting, this paper proposes a personalized federated decomposition-ensemble approach (PFDEA) for electricity load forecasting. First, a horizontal federated learning framework is constructed to enable secure data sharing and collaborative model training among different power companies. Second, a distributed data decomposition method is designed to decompose the original load sequences into multiple unified subcomponents, thereby improving the learning efficiency of local models. Finally, a personalized global ensemble method is proposed, allowing the trained model to capture global dominant trends while adapting to local data heterogeneity, and mitigating the communication overhead introduced by data decomposition. To validate the effectiveness of the proposed collaborative forecasting model, an empirical analysis is conducted using the GEFC2012 dataset, which contains real load data from 20 substations. The results demonstrate that the proposed method performs well in leveraging multi-source load data, addressing data distribution discrepancies, and enhancing communication efficiency.

Key words: Electric load forecasting, Personalized federated learning, Data sharing, Decomposition ensemble