Privacy-preserving probabilistic wind power forecasting: An adaptive federated approach

Accurate wind power forecasting (WPF) is crucial for the reliability of the power system operation and control. In recent years, probabilistic WPF has gained growing attention, and various advanced data-driven approaches have been proposed to achieve accurate probabilistic WPF. However, the data-driven approach relies on high-quality/volume data, which is hard to collect in reality, leading to the performance of these approaches falling short of expectations. This work proposes a federated learning (FL) based probabilistic WPF framework to utilize the data from other wind farms (WFs) to construct forecasting models while preserving privacy. To overcome the issue of non-independent and identically distributed data, an adaptive clustering strategy and elastic weight consolidation-based personalization have been proposed. The adaptive clustering strategy is adopted to separate the WFs into different clusters in the process of FL training. Additionally, elastic weight consolidation is introduced into the global model personalization process to prevent catastrophic forgetting. The experiments have been conducted with a dataset consisting of seven WFs across five forecasting settings. The results show that the proposed approach can achieve stable clustering convergence, higher accuracy, and more robust probabilistic WPF performance without the leakage of local data of WFs.