A component-aware progressive transfer learning method for wind power forecasting in newly built wind farms

Accurate wind power forecasting is essential for wind farm operation and grid integration. However, newly built wind farms often have limited historical data, restricting the performance of data-driven models. Under data scarcity, conventional deep learning models tend to overfit, while direct transfer learning may introduce irrelevant source-domain information when source and target farms differ. To address these challenges, this paper proposes a component-aware progressive transfer learning method for newly built wind farms. Relevant source wind farms are first selected based on combined power-sequence and meteorological similarity to reduce negative transfer. The original power sequence is then decomposed into trend and fluctuation components, which are modeled using a multilayer perceptron and a conditional generative adversarial network, respectively, and integrated through attention-based fusion. Rather than fine-tuning the entire model simultaneously, the proposed strategy progressively adapts the trend predictor, fluctuation model, and fusion module. Experiments on three newly built wind farm datasets yielded root mean square errors of 0.8850, 2.5060, and 5.0429 MW, representing reductions of 35.29%, 25.71%, and 33.03% over the best-performing benchmarks. On a United States dataset, the method achieved a root mean square error of 0.6543 MW and a coefficient of determination of 0.9870 for 15-min-ahead forecasting, while retaining its advantage over the comparison methods for forecasts 30 and 60 min ahead. These results demonstrate the effectiveness of source relevance selection, component-specific modeling, and progressive transfer learning under limited target-domain data.