【摘要】 针对风力发电功率序列的波动特性与非平稳特性引起的预测精度偏低问题，提出一种基于辛几何模态分解(symplectic geometry mode decomposition, SGMD)、Transformer架构与双向长短期记忆(bidirectional long short-term memory, BiLSTM)网络的超短期风力发电功率组合预测模型。首先，运用SGMD技术对初始风电功率数据进行自适应分解，有效分离高频干扰与低频趋势成分，并将其作为预测模型的输入特征；其次，针对分解后的子序列特性，利用Transformer的多头注意力机制捕捉多分量间的全局关联性，结合BiLSTM的双向时序建模能力增强局部特征提取；最终，通过整合各子序列预测结果，重构最终功率值。仿真结果表明，所提出的模型大幅提高了超短期风电功率预测的精度与稳定性，为高占比风电电力系统的安全稳定运行提供了可靠技术支撑。更多还原

【Abstract】 To address low forecasting accuracy caused by fluctuations and non-stationarity in wind power sequences, this study proposes a hybrid forecasting model integrating symplectic geometry mode decomposition(SGMD), Transformer architecture, and bidirectional long short-term memory(BiLSTM) networks. SGMD adaptively decomposes raw wind power data into distinct components, effectively separating high-frequency noise from low-frequency trends as model inputs. The Transformer-BiLSTM framework leverages multi-head attention to capture global interdependencies among components while utilizing bidirectional temporal modeling for local feature extraction. Final forecasts are reconstructed through component integration. Simulations demonstrate significantly enhanced accuracy and stability in ultra-short-term forecasting, providing reliable technical support for the secure operation of wind-dominant power systems.更多还原