作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (9): 286-297,304. doi: 10.19678/j.issn.1000-3428.0062059

• 开发研究与工程应用 • 上一篇    下一篇

结合注意力机制与LSTM的短期风电功率预测模型

廖雪超1,2, 伍杰平1,2, 陈才圣1,2   

  1. 1. 武汉科技大学 计算机科学与技术学院, 武汉 430065;
    2. 武汉科技大学 智能信息处理与实时工业系统重点实验室, 武汉 430065
  • 收稿日期:2021-07-13 修回日期:2021-10-14 发布日期:2022-09-08
  • 作者简介:廖雪超(1979—),男,副教授、硕士,主研方向为控制理论与控制工程;伍杰平、陈才圣,硕士研究生。
  • 基金资助:
    国家自然科学基金(61902285)。

Short-Term Wind Power Prediction Model Combining Attention Mechanism and LSTM

LIAO Xuechao1,2, WU Jieping1,2, CHEN Caisheng1,2   

  1. 1. School of Computer Science and Technology, Wuhan University of Science and Technology, Wuhan 430065, China;
    2. Key Laboratory of Intelligent Information Processing and Real-Time Industrial Systems, Wuhan University of Science and Technology, Wuhan 430065, China
  • Received:2021-07-13 Revised:2021-10-14 Published:2022-09-08

摘要: 风力发电预测在电力系统的运行中发挥着重要作用。现有风电功率的短期预测模型因风速的复杂性和随机性,难以确定风速与风电功率的非线性映射关系,导致预测精度降低。提出一种结合变分模态分解、双阶段注意力机制、误差修正模块与深度学习算法的短期风电功率预测模型。通过对原始数据进行互信息特征选择,获得与风电功率相关性较强的特征,并对其进行信号预处理,利用变分模态分解对多维特征序列进行分解,得到具有一定中心频率的模态分量,以降低各个特征序列的复杂性和非平稳性。采用基于双阶段注意力机制与编解码架构的长短时记忆(LSTM)神经网络对模态分量进行训练与预测,得到初始预测误差。在此基础上,利用误差修正模块对初始预测误差进行变分模态分解和修正,从而提高模型的预测精度。实验结果表明,与自回归移动平均模型、标准编解码结构的LSTM模型相比,该预测模型的平均绝对误差最高可降低约87%,具有较优的预测性能。

关键词: 短期风电功率预测, 变分模态分解, 长短时记忆神经网络, 注意力机制, 误差修正

Abstract: Wind power prediction plays an important role in operating power systems.Owing to the complexity and randomness of wind speed, determining the nonlinear mapping relationship between wind speed and power is difficult in existing short-term wind power prediction models, resulting in a decrease in prediction accuracy.This study proposes a short-term wind power prediction model that combines Variational Modal Decomposition (VMD), a dual-stage attention mechanism, an error correction module, and a deep learning algorithm.Through the Mutual Information(MI) feature selection of the original data, features with strong correlations with the wind power of the target prediction sequence are selected and the signal is preprocessed.The multi-dimensional features are decomposed using VMD to obtain modal components with a certain central frequency, thereby reducing the complexity and nonstationarity of each feature sequence.The Long Short-Term Memory (LSTM) neural network based on the dual-stage attention mechanism and encoder-decoder architecture is used to train and predict the modal components, and the initial prediction sequence is obtained.Further, the error correction module is used to perform VMD and error correction on the initial prediction error to improve the prediction accuracy of the model.The experimental results show that, compared with the autoregressive moving average model, LSTM model with a standard encoder-decoder structure, the average absolute error of the prediction model can be reduced by approximately 87% at its highest, and the prediction performance is optimal.

Key words: short-term wind power prediction, Variational Modal Decomposition(VMD), Long Short-Term Memory(LSTM) neural network, attention mechanism, error correction

中图分类号: