This paper presents a novel Federated Learning (FL) framework for day-ahead wind power forecasting in Taiwan, addressing data privacy and decentralized infrastructure challenges. Conventional methods struggle with data aggregation due to privacy concerns, and isolated local models lack generalizability. The proposed FL framework enables collaborative learning by training a shared global model without exchanging raw data among wind farms and weather stations. Each client uses a local Long Short-term Memory (LSTM)-based deep learning model to capture temporal dependencies in meteorological data. The framework employs a weighted federated averaging algorithm, prioritizing local model updates based on their forecasting performance. Experimental results from diverse Taiwanese locations show the FL model consistently outperforms independent local models in forecasting accuracy and robustness. The global model reduced the Root Mean Squared Error (RMSE) for Guanyin from 0.0985 to 0.0090, for Taipei from 0.0925 to 0.0167, and for Yangmingshan from 0.8629 to 0.1455, compared to individual local models.
Hong et al. (Mon,) studied this question.