面向光伏用户电压质量指标的注意力机制双流LSTM-CNN 预测方法

贺星; 王智; 刘谋海; 刘洁; 黄瑞; 颜宏文; 颜勤; 马瑞

贺星,王智,刘谋海,刘洁,黄瑞,颜宏文,颜勤,马瑞.面向光伏用户电压质量指标的注意力机制双流LSTM-CNN 预测方法[J].电测与仪表,2026,63(3):126-136.
Hexing,Wang Zhi,Liu Mouhai,Liu Jie,Hunag Ruin,Yan Hongwen,Yan Qin,Ma Rui.Attention mechanism dual-stream LSTM-CNN prediction model for photovoltaic user voltage quality index[J].Electrical Measurement & Instrumentation,2026,63(3):126-136.

面向光伏用户电压质量指标的注意力机制双流LSTM-CNN 预测方法

Attention mechanism dual-stream LSTM-CNN prediction model for photovoltaic user voltage quality index

DOI：10.19753/j.issn1001-1390.2026.03.013

中文关键词: 低压分布式发电光伏电站电压质量预测神经网络注意力机制特征融合

英文关键词:low-voltage distributed generation, photovoltaic power station, voltage quality prediction, neural network, attention mechanism, feature fusion

基金项目:国网湖南省电力有限公司科技项目(5216AG220007）

作者中文名	作者英文名	单位
贺星	Hexing	国网湖南供电服务中心 (计量中心)
王智	Wang Zhi	国网湖南供电服务中心 (计量中心)
刘谋海	Liu Mouhai	国网湖南供电服务中心 (计量中心)
刘洁	Liu Jie	长沙理工大学计算机与通信学院
黄瑞	Hunag Ruin	国网湖南供电服务中心(计量中心)
颜宏文	Yan Hongwen	长沙理工大学计算机与通信学院
颜勤	Yan Qin	长沙理工大学电气与信息工程学院
马瑞	Ma Rui	长沙理工大学电气与信息工程学院

摘要点击次数: 413

中文摘要:

随着国家双碳战略目标的提出,分布式光伏迎来新的发展机遇,对光伏发电站的电压质量管理成为关键挑战。文中基于双流LSTM-CNN(long short-term memory-convolutional neural network)模型,建立交互层和多特征融合预测模块来实现对光伏用户电压质量指标的预测,监控各设备在不同时段、相同间隔的运行状态反馈数据,对数据进行指数移动平均滤波处理来减少噪声的影响,构建双流LSTM-CNN 模型进行序列建模,并引入注意力机制增强对关键特征的关注程度,最后提出了多特征融合模块,充分利用来自不同层次的数字信息来丰富结果预测的特征表示,分别得到6 个预测头来实现对未来电压闪变、电压波动、电压偏差(电压质量指标) 的预测。实验使用平均绝对误差(mean absolute error, MAE) 和均方根误差(root mean square error, RMSE)对模型性能进行评估,其中电压偏差MAE 的仿真结果为0. 029 0,表现出较小的预测误差。实验结果表明,该方法能够有效地预测低压分布式光伏电压质量的情况,为低压分布式光伏台区的稳定运行提供重要支持和参考。

英文摘要:

With the proposal of the national dual-carbon strategic goal, distributed photovoltaic (PV) ushers in new development opportunities, and the management of voltage quality for photovoltaic power stations becomes a key challenge. Based on the dual flow long short-term memory-convolutional neural network (LSTM-CNN) model, this paper proposes an interaction layer and a multi-feature fusion prediction module to achieve the prediction of photovoltaic user voltage quality indicators. The operating status feedback data of different periods and the same interval is monitored, and the data is filtered by exponential moving average to reduce the influence of noise. The dual-stream LSTM-CNN model for sequence modeling is constructed, with the introduction of attention mechanism aimed at enhancing attention to key features. We propose a multi-feature fusion module, which fully leverages digital information from different levels to enrich the feature representation for prediction. This module consists of six prediction heads, allowing us to predict future voltage fluctuations, voltage variations, and voltage deviations (voltage quality indices). Model performance is evaluated using the mean absolute error (MAE) and the root mean square error (RMSE), where the simulation of voltage deviation MAE is 0. 029 0, showing a small prediction error. The experimental results demonstrate that the proposed method can effectively predict the voltage quality of low voltage distributed PV, and provide important support and reference for the stable operation of low voltage distributed PV platform area.

查看全文查看/发表评论下载PDF阅读器

关闭