Abstract: The data dependence of deep learning models for wind power forecasting presents a significant challenge for newly built wind farms, which may lack sufficient data for effective model training. Overcoming this limitation by amalgamating data from other wind turbines introduces an inherent risk of privacy leakage. To address this issue, we introduced a novel hybrid model for wind power forecasting, specifically designed to address data privacy preservation. This model integrates a temporal convolutional network with a transformer encoder, harnessing the strengths of both components. The temporal convolutional network extracts local temporal patterns, whereas the transformer encoder captures the intricate time dependencies between these patterns. To address the data privacy risks caused by traditional centralized forecasting methods, a federated training strategy is implemented to train a collaborative knowledge-sharing model, ensuring that the original data from the source domains remain confidential and undisclosed. The experimental results validate the effectiveness of the proposed model, showing significant improvements in forecasting performance. When federated with other wind turbines, the average mean absolute error and the root mean square error of six wind turbines are 4.624 6 and 6.245 7, respectively. In contrast to traditional centralized training methods, the proposed federated learning method, which was applied to inland wind turbines, achieved almost no increase in losses.

The full text can be downloaded at https://doi.org/10.1007/s40436-025-00552-1

Key words: Temporal convolutional network, Privacy-preserving, Federated learning, Wind power forecasting, Transformer