摘要
为有效分析与利用光伏功率预测模型中以特定规律分布的预测误差,提出基于LSTM-Attention和CNN-BiGRU误差修正的光伏功率预测模型。首先,引入注意力机制(Attention)弥补输入序列长时长短期记忆网络(LSTM)难以保留关键信息的不足,建立LSTM-Attention的预测模型对光伏功率进行初步预测。其次,将卷积神经网络(CNN)在非线性特征提取上的优势与双向门控循环单元(BiGRU)在防止多种特征相互干扰的优势相结合,搭建CNN-BiGRU误差预测模型对可能产生的误差进行预测,从而对初步预测结果进行修正。经过实例分析表明:与未经误差修正的预测结果进行对比,经CNN-BiGRU误差预测模型进行误差修正后在不同天气类型中均能有效提高预测精度。
To effectively analyze and utilize the prediction errors distributed in a specific pattern in the photovoltaic power prediction model,a photovoltaic power prediction model based on LSTM-Attention and CNN-BiGRU error correction is proposed.Firstly,the LSTM-Attention mechanism is introduced to compensate for the shortcomings of the long short-term memory(LSTM)network,which is difficult to retain key information in the input sequence.Secondly,the advantages of convolutional neural network(CNN)in non-linear feature extraction are combined with the advantages of Bidirectional gated recurrent unit(BiGRU)in preventing multiple features from interfering with each other to build a CNN-BiGRU error prediction model is used to predict the possible errors and to correct the initial prediction results.The experimental results indicates that the CNN-BiGRU error prediction model can effectively improve the prediction accuracy in different weather types when compared with the prediction results without error correction.
作者
吐松江·卡日
雷柯松
马小晶
吴现
余凯峰
Kari·Tusongjiang;Lei Kesong;Ma Xiaojing;Wu Xian;Yu Kaifeng(School of Electrical Engineering,Xinjiang University,Urumqi 830049,China;State Grid Changji Power Supply Company,Changji 831100,China)
出处
《太阳能学报》
CSCD
北大核心
2024年第12期85-93,共9页
Acta Energiae Solaris Sinica
基金
新疆维吾尔自治区自然科学基金(2022D01C35)
国家自然科学基金(52067021)
新疆维吾尔自治区优秀青年科技人才培养项目(2019Q012)。
关键词
光伏功率预测
深度学习
误差修正
注意力机制
长短期神经网络
双向门控循环单元
photovoltaic power prediction
deep learning
error correction
attention mechanism
long-short term memory network
bidirectional gated recurrent unit
作者简介
通讯作者:吐松江·卡日(1984—),男,博士、副教授、博士生导师,主要从事电力系统数字化、人工智能与模式识别等方面的研究。minyun229@163.com。
