A Daily Load Forecasting Method Based on Federated Learning and Transformer Model

In view of the increasing difficulty of regional short‐term load forecasting and user data protection under the complex environment of energy Internet, a daily load forecasting method based on federated learning and transformer model is proposed. Firstly, design a dynamic privacy budget to improve the differential privacy algorithm, and introduce it into federated learning for privacy protection of user electricity meter data collection. Then, combining the improved graph convolutional network‐time convolutional network (GCN‐TCN) and transformer models to analyse the load data, and using it as a local model for federated learning for collaborative training. Finally, establish a daily load forecasting framework based on federated learning, aggregate model parameters through the FedAdam algorithm and construct a global model to achieve high‐precision load forecasting. Based on the selected dataset, experimental analysis is conducted on the proposed method, and the results show that its predicted results are very close to the actual values. Taking workdays as an example, its mean absolute percentage error, root mean square logarithmic error and mean absolute error are 1.573%, 0.019 and 23.581 MW, respectively, and could effectively protect privacy and predictive performance.