基于双频编码器-解码器架构的超短期多步风速预测模型

Ultra-short-term multistep wind speed prediction model based on dual frequency encoder-decoder

摘要

【目的】针对风速非平稳特性与深度学习模型固有局限导致的预测精度提升困难制约风力发电高效利用的关键瓶颈，探索高精度超短期风速预测技术。【方法】以江西两座风场为研究对象，构建基于注意力机制和长短期记忆网络(LSTM)的双频率编码器-解码器模型。该模型应用卷积理论对风速序列进行多尺度分解，有效缓解非平稳性影响；创新设计频时注意力机制，通过频域-时域特征交互，自适应捕捉高频波动特征；采用两层LSTM的编码器-解码器架构，利用层级非线性映射深度挖掘低频趋势特征，并抑制多步预测中的误差累积。【结果】基于两座风场数据的对比实验显示，该模型在15步、30步和60步预测中均表现优异：30步预测平均绝对误差(MAE)为0.340 1、均方误差(MSE)为0.281 6、均方根误差(RMSE)为0.530 6;60步预测MAE为0.471 3、MSE为0.504 7、RMSE为0.710 2,较对比模型预测精度提升显著，且泛化能力突出。【结论】该模型通过频域-时域协同建模与层级特征挖掘，有效解决风速非平稳性和模型精度瓶颈问题，在多步预测中展现出高精度和优异泛化性，为风力发电场景下的超短期风速预测提供了可靠的技术方案。

Abstract

[Objective] The high-precision ultra-short-term wind speed prediction technology is investigated to address the key bottleneck constraining the efficient utilization of wind power: the challenge of improving prediction accuracy caused by the non-stationary nature of wind speed and the inherent limitations of deep learning models.[Methods] Taking two wind farms in Jiangxi as the research objects, a dual-frequency encoder-decoder model based on an attention mechanism and long short-term memory(LSTM) was established. The model applied convolution theory to perform multi-scale decomposition of wind speed sequences, effectively mitigating the impact of non-stationarity. A novel frequency-time attention mechanism was proposed to adaptively capture high-frequency fluctuation characteristics through the interaction of features in the frequency and time domains. A two-layer LSTM encoder-decoder architecture was adopted to effectively extract low-frequency trend features through hierarchical nonlinear mapping and suppress error accumulation in multistep predictions.[Results] Comparative experiments using data from the two wind farms showed that the proposed model demonstrated excellent performance in 15-step, 30-step, and 60-step predictions. For 30-step prediction, the mean absolute error(MAE), mean squared error(MSE), and root mean squared error(RMSE) were 0.340 1, 0.281 6, and 0.530 6, respectively. For 60-step prediction, the MAE, MSE, and RMSE were 0.471 3, 0.504 7, and 0.710 2, respectively. This demonstrated that the proposed model showed a significant improvement in prediction accuracy compared to baseline models, along with outstanding generalization ability.[Conclusion] By leveraging collaborative modeling in the frequency and time domains and hierarchical feature extraction, the proposed model effectively addresses the non-stationarity of wind speed and the limitations in model accuracy. It demonstrates high precision and excellent generalization in multistep prediction, providing a reliable technical solution for ultra-short-term wind speed prediction in wind power generation scenarios.

关键词

深度学习 / Transformer / 频时注意力机制 / 双层长短期记忆网络 / 解码器-编码器 / 多步预测 / 气候变化 / 影响因素

Key words

deep learning / Transformer / frequency-time attention mechanism / two-layer long short-term memory units / encoder-decoder / multistep prediction / climate change / influencing factors

基金资助

国家自然科学基金项目(52309111)

中国水利水电科学研究院基本科研业务费项目“智慧水利水电工程一体化平台技术研究”(AU0145B022021)