Wind power forecasting for newly built wind farms based on deep learning with dual-stage attention mechanism and adaptive transfer learning

Accurate wind power forecasting (WPF) is critical for optimal wind power scheduling. While deep learning methods are effective for WPF modeling, they struggle with insufficient data for newly built wind farms. To address this issue, we propose a long short-term memory network (LSTM) with dual-stage attention and adaptive transfer learning (LSTM-DSA-ATL), comprising two stages: multi-source pre-training and adaptive transfer learning. In the pre-training stage, multi-source wind farm data are selected using the maximum information coefficient (MIC) criterion. A LSTM with dual-stage attention (LSTM-DSA) is then developed for pre-training base model to extract spatio-temporal features from large-scale historical data of multiple source wind farms. During the transfer learning stage, an evolutionary optimization based adaptive transfer learning strategy is proposed, transforming the transfer learning problem into an optimization task with a novel accuracy-stability objective. This enables dynamic adaptation to newly built wind farms through adaptive selection of transfer content and strategies across multi-source farms, while preserving rich multi-source knowledge to ensure prediction stability. The proposed method significantly enhances the accuracy, reliability, and generalization capability of WPF models. Its effectiveness is validated using two real-world wind power datasets, demonstrating superior performance over traditional methods in forecasting accuracy, predictive power, and generalization.