A privacy-preserving framework integrating federated learning and transfer learning for wind power forecasting

Data-driven approaches show significant potential in accurately forecasting the power generation of wind turbines. However, it suffers from a lack of training data in various scenarios. Transfer learning has been employed as a solution to address the issue of limited data by leveraging data from other turbines. However, the conventional centralized training approach raises concerns about data privacy risks. In this study, we propose a privacy-preserving framework that incorporates federated learning and transfer learning for wind power forecasting. Firstly, we design a hybrid model with a series architecture as the backbone model to improve forecasting accuracy. Secondly, the proposed two-stage framework consists of federated learning-based pre-training and personalized fine-tuning. The federated learning stage pre-trains a knowledge-sharing model without disclosing raw data from source domains. Based on the pre-trained global model, personalized fine-tuning is applied to establish a customized model for the target turbine. Experimental results demonstrate that the proposed hybrid model achieves better forecasting accuracy. Additionally, the two-stage framework not only addresses insufficient data while considering privacy preservation but also enhances personalized adaptation of shared knowledge for the target turbine. Compared to local training and traditional federated learning, the proposed framework demonstrates obvious accuracy improvements, reaching up to 43.32 % and 27.94 %, respectively.