摘要
新型电力系统下新能源在大规模并网的同时亦带来诸多的电能质量问题,严重影响电网的安全稳定运行。准确识别电能质量扰动(PQDs)类型是保障用电设备稳定运行、实现电力高质量供应的重要前提。针对现有PQDs识别和分类方法普遍存在的模型识别准确率低、参数量大、算法复杂度较高等问题,该文提出一种基于条件网络和知识注入的多标签学习PQDs识别与分类方法。该方法首先在主网络中采用一维融合卷积网络与基于无参注意力机制优化的长短期记忆网络来提取PQDs特征信息。然后,设计了一个条件网络用于识别PQDs扰动数量并引入知识注入模块增强模型的约束能力,通过结合条件网络和知识注入机制,在减少主网络对复杂扰动类型识别难度的同时提高了该文算法识别的准确度。最后,算法将条件网络的分类结果和主网络的输出结果通过一个标签阈值函数得到最终扰动识别与分类结果。仿真结果表明,在随机白噪声环境下,该文所提方法以仅92654的参数量达到了99.47%的识别准确率,实现了复杂扰动的高精度和轻量化识别。此外,实际构建的硬件平台进一步验证了所提方法的准确性和可靠性。
The large-scale integration of renewable energy into the new power system has brought numerous power quality disturbances(PQDs),which severely affect the safe and stable operation of the power grid.Accurate identification of PQD types is crucial for ensuring the stable operation of electrical equipment and achieving high-quality power supply.To address the common issues of low recognition accuracy,large model parameters,and high computational complexity in existing PQDs recognition and classification methods,this paper proposes a multi-label learning method for PQDs identification and classification based on conditional networks and knowledge injection.Firstly,this method integrates a one-dimensional convolutional neural network(1D-CNN)with a one-dimensional group convolutional neural network(1D-GCCN)to construct a one-dimensional multi-fusion convolutional network(1D-MFCN),which effectively extracts key features from PQDs signals while reducing the computational complexity of the model.Subsequently,to further capture the temporal characteristics of PQDs,a long short-term memory(LSTM)network optimized by a simple,parameter-free attention mechanism(SimLSTM)is proposed.Compared with traditional LSTM,the SimLSTM is capable of automatically identifying salient features and assigning appropriate weights,thereby enhancing the precision of temporal feature extraction.In addition,a conditional network is designed to predict the number of PQDs occurring simultaneously,and a knowledge injection module is constructed by leveraging the mutually exclusive nature of swell,sag,and interruption disturbances.By incorporating the conditional network and knowledge injection mechanism,the proposed method alleviates the burden on the main network in distinguishing complex disturbance types and improves the overall classification accuracy.Finally,the outputs from the conditional network and the main network are fused through a label threshold function to produce the final disturbance recognition and classification results.Simulation results demonstrate that,under the presence of random white noise,the proposed method achieves an accuracy of 99.47%with only 0.093 M parameters,enabling high-precision and lightweight recognition of complex disturbances.Furthermore,when tested on 14 types of PQDs collected from an actual hardware platform,the proposed model attains an average recognition accuracy of 98.45%,with a perfect 100%accuracy on four-type disturbances,further validating the reliability and effectiveness of the proposed approach.
作者
黄杰
李建闽
朱冰凡
张驰成
梁成斌
Huang Jie;Li Jianmin;Zhu Bingfan;Zhang Chicheng;Liang Chengbin(College of Engineering and Design Hunan Normal University,Changsha 41008,China;College of Electrical and Information Engineering Hunan University,Changsha 410082 China;College of Electrical Engineering Guizhou University,Guiyang 550025 China)
出处
《电工技术学报》
北大核心
2025年第23期7652-7663,共12页
Transactions of China Electrotechnical Society
基金
国家自然科学基金重点项目(51907062)
湖南省自然科学基金(2021JJ40354)
贵州省科技支撑计划[2024]资助。
关键词
电能质量扰动
条件网络
知识注入
多标签学习
长短期记忆网络
Power quality disturbances
conditional network
knowledge injection
multi-label learning
long short-term memory(LSTM)
