摘要
新能源背景下,电能质量扰动(power quality disturbances,PQDs)常面临扰动信息利用率低和易受噪声干扰等问题。为此,提出一种基于通道多头自注意力机制优化的残差密集连接空洞空间金字塔池化层(residual densely connected atrous spatial pyramid pooling optimized by channel multi-head self-attention mechanism,RDASPP-CMSA)用于PQDs分类。使用残差密集连接空洞空间金字塔池化层(residual densely connected atrous spatial pyramid pooling,RDASPP)提取和融合目标的多尺度全局特征,引入残差连接和BottleNeck模块显著提高模型的感知能力和稳定性。结合全局池化和多头自注意力机制,提出一种动态加权的特征选择策略,即通道多头自注意力机制(channel multi-head self-attention mechanism,CMSA);通过在通道和时间维度上同时进行特征学习,动态加权选择特征,识别扰动中的关键特征;通过全连接层精准实现对每个扰动信号的分类。仿真结果显示:RDASPP-CMSA在30 dB噪声环境下对29种扰动的识别准确率为99.60%,对5种实际电网扰动识别准确率达到99.98%,在分类准确率和抗噪性能方面均表现良好,具有分类准确率高、噪声鲁棒性强等优点。
Power quality disturbances(PQDs)often encounter such problems as low utilization of disturbance information and vulnerability to noise interference.To address these issues,a new model,Residual Densely Connected Atrous Spatial Pyramid Pooling Optimized by Channel Multi-Head Self-Attention Mechanism(RDASPP-CMSA)is built.It is designed specifically for PQD classification.First,the Residual Densely Connected Atrous Spatial Pyramid Pooling(RDASPP)module is employed to extract and integrate multi-scale global features.By incorporating residual connections and BottleNeck structures,the model improves stability and enhances feature perception capabilities.Then,a dynamic weighting feature selection strategy,Channel Multi-Head Self-Attention Mechanism(CMSA),is introduced through the integration of global pooling and multi-head self-attention mechanisms.It facilitates simultaneous feature learning across both channel and temporal dimensions,achieving dynamic weighting and selection of features.Thus,the model effectively suppresses noise and captures critical features within disturbance signals.Finally,a fully connected layer is utilized to perform precise classification of individual disturbance signals.Experimental results reveal the RDASPP-CMSA model achieves a classification accuracy of 99.60%for 29 types of disturbances under a 30 dB noise environment and an accuracy of 99.98%when tested on five real-world grid disturbances.Comparative analysis shows it outperforms existing models in both classification accuracy and robustness against noise,underscoring its efficacy in PQD classification tasks.
作者
叶鹏
宋弘
吴浩
江俊卓
YE Peng;SONG Hong;WU Hao;JIANG Junzhuo(School of Automation and Information Engineering,Sichuan University of Science and Engineering,Zigong 643000,China;Key Laboratory of Artificial Intelligence in Sichuan Province,Zigong 643000,China;Aba Teachers University,Aba 623002,China)
出处
《重庆理工大学学报(自然科学)》
北大核心
2025年第5期202-210,共9页
Journal of Chongqing University of Technology:Natural Science
基金
四川省科技厅项目(2022YFS0518,2022ZHCG0035)
人工智能四川省重点实验室项目(2022RZY01)。
关键词
电网
电能质量
注意力机制
深度学习
power grid
power quality
attention mechanism
deep learning
