The efficient utilization of wind energy relies on accurate wind power forecasting. However, existing methods face challenges in multi-step forecasting, including error accumulation, insufficient utilization of numerical weather prediction (NWP), and inadequate modeling of localized meteorological characteristics. To address these issues, this paper proposes a heterogeneous sequence-to-sequence (seq2seq) model based on PatchTST-GRU, integrated with NWP refinement. The heterogeneous seq2seq architecture employs a PatchTST backbone as the encoder to extract local temporal features from historical wind power data, combined with a GRU decoder to generate prediction sequences, thereby enhancing long-term dependency modeling. The NWP refinement module is developed to improve the usability of low-resolution NWP data through spatiotemporal scaling, providing future meteorological trend guidance for the decoder. Furthermore, a fusion attention mechanism with asymmetric query-key-value matrices is introduced to dynamically fuse historical wind power temporal patterns with refined NWP features, optimizing prediction outcomes. Experiments using real-world wind farm data demonstrate the effectiveness of the heterogeneous seq2seq architecture, NWP refinement and fusion attention mechanism. Within 48 to 288 forecasting steps, the proposed method outperforms conventional approaches, including light gradient boosting machine, support vector regression and seq2seq, in multiple evaluation metrics. This framework provides reliable multi-step power forecasting support for wind farm operations, while its heterogeneous seq2seq architecture exhibits potential for application to other time series prediction tasks.
Keywords: NWP refinement; PatchTST; Seq2seq; Wind power forecasting.
© 2025. The Author(s).