发电技术 ›› 2025, Vol. 46 ›› Issue (1): 93-102.DOI: 10.12096/j.2096-4528.pgt.23101

• 新能源 • 上一篇    下一篇

基于多特征提取-卷积神经网络-长短期记忆网络的短期风电功率预测方法

匡洪海, 郭茜   

  1. 湖南工业大学电气与信息工程学院,湖南省 株洲市 412007
  • 收稿日期:2023-08-25 修回日期:2023-12-18 出版日期:2025-02-28 发布日期:2025-02-27
  • 作者简介:匡洪海(1972),女,博士,教授,研究方向为新能源发电与并网,khhzyz@163.com
    郭茜(1996),女,硕士研究生,研究方向为风电功率预测,12458044@qq.com
  • 基金资助:
    国家自然科学基金项目(51977072)

Short-Term Wind Power Prediction Method Based on Multimodal Feature Extraction-Convolutional Neural Network-Long-Short Term Memory Network

Honghai KUANG, Qian GUO   

  1. College of Electrical and Information Engineering, Hunan University of Technology, Zhuzhou 412007, Hunan Province, China
  • Received:2023-08-25 Revised:2023-12-18 Published:2025-02-28 Online:2025-02-27
  • Supported by:
    National Natural Science Foundation of China(51977072)

摘要:

目的 天气和随机因素会改变误差的统计特征,因此考虑对影响风电功率的多种气候因素进行特征提取,为优化功率时序特征提取,提出基于多特征提取(multimodal feature extraction,MFE)-卷积神经网络(convolutional neural network,CNN)-长短期记忆(long-short term memory,LSTM)网络的风电功率预测方法。 方法 首先,对数值天气预报(numerical weather prediction,NWP)数据提取11种统计性特征,通过提取基本特征和统计性特征对原始数据进行聚类,并根据类别分别建立预测模型,以提高预测模型的适应性;其次,在网络架构上对LSTM进行改进,通过CNN的特征提取能力和LSTM的非线性序列预测能力,实现对风电功率历史信息和NWP数据的充分挖掘。最后,利用我国新疆某风电场数据,通过MFE消融实验、CNN消融实验,验证了所提短期风电功率预测方法的有效性和优越性。 结果 相比于自回归移动平均(autoregressive integrated moving average,ARIMA)、全连接循环神经网络(fully recurrent neural network,FRNN)模型和MFE-LSTM、CNN-LSTM模型,MFE-CNN-LSTM预测方法的均方根误差与平均绝对误差均有所下降。 结论 MFE-CNN-LSTM预测方法可有效提取特征,并且MFE与CNN有效提升了预测准确性。

关键词: 多特征提取, 卷积神经网络, 长短期记忆网络, k-均值聚类算法, 风电功率预测, 短期预测, 消融实验

Abstract:

Objectives Weather and random factors can alter the statistical characteristics of errors. Therefore, this study considers feature extraction of various climate factors that affect wind power. To optimize the extraction of power time series features, a wind power prediction method based on multi-feature extraction (MFE), convolutional neural network (CNN), and long short-term memory (LSTM) network is proposed. Methods Firstly, 11 statistical features are extracted from numerical weather prediction (NWP) data. By extracting basic and statistical features, the original data is clustered, and prediction models are established according to categories to improve the adaptability of prediction models. Next, the network architecture of LSTM is improved. By leveraging the feature extraction ability of CNN and the nonlinear sequence prediction ability of LSTM, the historical information of wind power and NWP data is thoroughly explored. Finally, using the data from a wind farm in Xinjiang, China, the effectiveness and advantages of the proposed short-term wind power prediction method are verified by MFE and CNN ablation experiments. Results The MFE-CNN-LSTM prediction method shows a decrease in both root mean square error and mean absolute error, compared with the autoregressive integrated moving average (ARIMA), fully recurrent neural network (FRNN), MFE-LSTM and CNN-LSTM models. Conclusions The MFE-CNN-LSTM prediction method can effectively extract features, and MFE and CNN effectively improve prediction accuracy.

Key words: multi-feature extraction, convolutional neural network, long short-term memory networks, k-means clustering algorithm, wind power prediction, short-term prediction, ablation experiment

中图分类号: