面向多尺度的多分支注意力风电功率预测模型
Multi-Scale Multi-Branch Attention Wind Power Forecasting Model
高子墨 1刘婧1
作者信息
- 1. 辽宁工程技术大学电子与信息工程学院，葫芦岛市 125105
- 折叠
摘要
针对风电功率具有强非线性、高不确定性和多尺度特征、传统统计与深度学习方法难以兼顾长时依赖与局部剧烈波动的问题，构建了一种多分支融合的 Informer-BiGRUGATT-CrossAttention 风电功率预测模型。模型采用 Informer 分支高效刻画跨日、跨周等长时间尺度的全局相关性，利用 BiGRU-Global Attention 分支提取突变段和高频波动等细粒度局部动态特征，并通过交叉注意力模块实现全局趋势与局部细节以及多源气象、机组状态信息的深度对齐与融合。基于某风电场 2019 年实测 15 min 分辨率多变量数据集开展多步预测实验，在 R2、NMAE、NRMSE 三项评价指标上均优于 CNN、TCN、BiTCN、Transformer 与标准 Informer 等基线模型，其中 R2 提升至 0.978，NRMSE 与 NMAE 分别降至 0.145 和 0.092，表明所构建模型在复杂风况和长时跨度场景下兼具较高预测精度与良好泛化能力，可为风电场优化运行及电网调度提供可靠的决策支撑。
Abstract
To address the challenges of strong nonlinearity, high uncertainty, and multi-scale characteristics in wind power, as well as the difficulty for traditional statistical and deep learning methods to simultaneously capture long-term dependencies and local drastic fluctuations, a multi-branch fused Informer-BiGRUGATT-CrossAttention wind power forecasting model was developed. The model uses the Informer branch to efficiently capture global correlations over long time scales such as day-to-day and week-to-week periods. The BiGRU-Global Attention branch extracts fine-grained local dynamic features such as abrupt changes and high-frequency fluctuations. Moreover, a cross-attention module is employed to achieve deep alignment and fusion of global trends, local details, and multi-source meteorological and turbine status information. Multi-step prediction experiments were conducted based on a 2019 measured 15-minute resolution multivariate dataset from a wind farm. The model outperformed baseline models including CNN, TCN, BiTCN, Transformer, and standard Informer in terms of R2, NMAE, and NRMSE metrics, achieving an R2 of 0.978, and reducing NRMSE and NMAE to 0.145 and 0.092, respectively. This demonstrates that the proposed model achieves high predictive accuracy and good generalization capability under complex wind conditions and long-term scenarios, providing reliable decision support for wind farm operation optimization and grid dispatching.关键词
电力系统及其自动化/风电功率预测/Informer 模型/BiGRU 神经网络/交叉注意力机制。Key words
Power systems and their automation/wind power forecasting/Informer model/BiGRU neural network/cross-attention mechanism引用本文复制引用
高子墨,刘婧.面向多尺度的多分支注意力风电功率预测模型[EB/OL].(2025-12-29)[2025-12-30].http://www.paper.edu.cn/releasepaper/content/202512-43.学科分类
风能、风力机械
评论