电工技术学报  2023, Vol. 38 Issue (12): 3277-3288    DOI: 10.19595/j.cnki.1000-6753.tces.220477
电力系统与综合能源 |
基于特征交叉机制和误差补偿的风力发电功率短期预测
刘雨佳, 樊艳芳, 白雪岩, 宋雨露, 郝瑞鑫
新疆大学电气工程学院 乌鲁木齐 830092
Short-Term Wind Power Prediction Based on Feature Crossover Mechanism and Error Compensation
Liu Yujia, Fan Yanfang, Bai Xueyan, Song Yulu, Hao Ruixin
School of Electrical Engineering Xinjiang University Urumchi 830092 China
全文: PDF (1478 KB)   HTML
输出: BibTeX | EndNote (RIS)      
摘要 为提高短期风电功率预测精度,首先在卷积神经网络(CNN)-长短期记忆(LSTM)网络模型的基础上,引入特征交叉(FC)机制,对风电场数据集进行相关性分析并交叉组合,增加特征维度,加强非线性特征学习,挖掘隐藏关联,提高训练精度,构建形成FC-CNN-LSTM预测模型;然后,将该预测模型在风电预测中产生的误差值作为训练数据,训练生成误差补偿模型,利用该模型计算结果对风电预测数据进行补偿,进一步提高预测精度;最后,通过仿真验证该方法具有较高的预测精度,且相比传统预测模型,在分钟级超短期尺度上的预测性能具有显著优势。
服务
把本文推荐给朋友
加入我的书架
加入引用管理器
E-mail Alert
RSS
作者相关文章
刘雨佳
樊艳芳
白雪岩
宋雨露
郝瑞鑫
关键词 卷积神经网络长短期记忆网络风功率预测特征交叉误差补偿    
Abstract:At present, wind power forecasting methods mainly include physical modeling, statistical modeling and artificial intelligence algorithm modeling. The traditional physical modeling and statistical modeling methods are difficult to collect data and select parameters, and have weak processing ability for a large number of data, so it is difficult to establish an accurate prediction model. Therefore, in practical applications, artificial intelligence algorithms are usually used to predict wind power. However, the current research on wind power prediction focuses on the use and improvement of artificial intelligence algorithms, and does not take into account the correlation between different features in the data and wind power, as well as the difference in the size of correlation. The model trained on this basis can only establish a single, superficial correlation, and cannotmine deeper relationships, which is not conducive to short-term prediction of wind power.
Therefore, this paper first introduces the feature crossover mechanism, analyzes the correlation of data features and cross combines them, increases feature dimensions, strengthens the learning ability of the algorithm for nonlinear features and deep hidden associations, and forms the FC-CNN-LSTM prediction model based on CNN-LSTM network improvement. Then, use the error value generated by the prediction model in the prediction as the training data, train and generate the error compensation model, and use the data generated by the error compensation model to compensate the wind power prediction data, so as to further improve the prediction accuracy. Finally, through the measured data of a wind farm, it is verified that the FC-CNN-LSTM model has a higher prediction accuracy, and after the error compensation process is added, it can further reduce the error compared with the traditional prediction methods, which has significant advantages.
The simulation results of the actual data in a region show that: (1) the feature crossing mechanism can effectively improve the learning ability of the model for nonlinear features and deep hidden associations, thus improving the prediction accuracy of the model. The prediction accuracy of the FC-CNN-LSTM model is 14.3% higher than that of the CNN-LSTM model; (2) The error compensation model based on FC-CNN-LSTM model can accurately predict the error of power prediction model, greatly reducing the prediction error, and greatly improving the prediction accuracy after compensation by 46.5% compared with the FC-CNN-LSTM model before compensation;
Finally, the following conclusions are drawn through analysis: (1) Compared with the CNN-LSTM model, the FC-CNN-LSTM model proposed in this paper has obvious advantages in the accuracy of the prediction of the ultra short termminute level wind power. It can more keenly capture the subtle changes of the characteristics related to the wind power, adapt to the rapid changes of meteorological factors, and more accurately predict the wind power, which is more suitable for practical engineering projects; (2) The error compensation mechanism based on FC-CNN-LSTM model proposed in this paper can further improve the accuracy of wind power prediction, and can be used in different application scenarios and with different algorithms, with high adaptability; (3) In addition, the disadvantage of the FC-CNN-LSTM model is that when the wind power is small and close to zero, the wind speed decreases significantly, and the contribution of other weak correlation features in the model will be relatively improved. Although these weak correlation features are related to the wind power to some extent, they do not occupy the same dominant position as the wind speed. Therefore, when the wind speed decreases significantly, the FC-CNN-LSTM model improves the learning of weak correlation features, The prediction accuracy will be worse than that of the CNN-LSTM model, which can be solved by setting a limit on the prediction results according to the wind speed.
Key wordsConvolution neural network    short and long term memory network    wind power prediction    feature crossover    error compensation   
收稿日期: 2022-04-01     
PACS: TM614  
  TM732  
基金资助:国家自然科学基金(51767023)和新疆维吾尔自治区研究生科研创新计划(XJ2022G041)资助项目
通讯作者: 樊艳芳 女,1971年生,教授,博士生导师,研究方向为新能源并网技术及电力系统保护与控制。E-mail:fyf3985@xju.edu.cn   
作者简介: 刘雨佳 男,1998年生,硕士研究生,研究方向电力系统控制与优化调度。E-mail:1104680459@qq.com
引用本文:   
刘雨佳, 樊艳芳, 白雪岩, 宋雨露, 郝瑞鑫. 基于特征交叉机制和误差补偿的风力发电功率短期预测[J]. 电工技术学报, 2023, 38(12): 3277-3288. Liu Yujia, Fan Yanfang, Bai Xueyan, Song Yulu, Hao Ruixin. Short-Term Wind Power Prediction Based on Feature Crossover Mechanism and Error Compensation. Transactions of China Electrotechnical Society, 2023, 38(12): 3277-3288.
链接本文:  
https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.220477          https://dgjsxb.ces-transaction.com/CN/Y2023/V38/I12/3277