A pattern-aware two-stage power forecasting method for offshore wind farms based on dynamic feature decoupling

Accurate offshore wind power forecasting (OWF) is challenged by strong non-stationarity and diverse weather-dependent operating patterns, leading to complex wind speed–power relationships and limiting the reliability of short-term predictions. To address these issues, a unified pattern-aware two-stage framework is proposed to jointly enhance feature representation and multi-step prediction stability within a single model. By incorporating pattern-conditioned information, operating patterns are embedded as conditional representations, and a modulation mechanism is introduced to guide feature learning for adaptive modeling across different regimes. A multi-scale conditional convolution module extracts multi-scale features from numerical weather prediction wind speed, while a dynamic feature decoupling mechanism separates the trend and seasonal components of wind power series, thereby improving the representation of complex temporal dynamics. In addition, a two-stage progressive training strategy alleviates the training–inference mismatch in recursive forecasting (RF), thereby mitigating error accumulation. Experiments on real-world offshore wind farm data under multiple forecasting strategies demonstrate that the proposed method consistently outperforms both classical and state-of-the-art models across all evaluation metrics. The method achieves an NRMSE of 12.25%, an NMAE of 9.03%, and a qualified rate of 94.83%, with a 7.1% relative reduction in NRMSE compared with the strongest baseline (ModernTCN with RF and scheduled sampling). The proposed framework provides an effective and practical solution for improving both accuracy and stability in day-ahead OWF.