Abstract: Some entities of power market lack sufficient data, leading to low accuracy in load forecasting. The traditional federated learning forecasting method can fully train a global model with multi-client data while preserving each client’s data privacy. However, the global model may not perform optimally on each client due to the neglect of personalized characteristics. Therefore, this paper proposes a personalized federated multi-region load forecasting method based on data decomposition (pFedD). First, the original load data sequence is decomposed to obtain intrinsic mode functions (IMF) containing different data features. Then, using the zero-crossing rate, all IMFs are divided into high-frequency components, low-frequency components, and trend components. Finally, through component correlations analysis, the high-frequency components and trend components are retained in the client as personalized components for local model training, and the low-frequency components participate as federated components of the central server for global model training. Experiments conducted on the load data of 10 areas in northern China demonstrate that the proposed method achieves an average mean absolute percentage error (MAPE) of 3.09%, representing a 1.67% reduction compared to the traditional federated averaging (FedAvg).