目的目的为了实现电器设备的长期稳定运行,方法方法提出基于长短时记忆(longshort-termmemory,LSTM)网络和多头注意力机制的气体绝缘开关柜(gasinsulatedswitchgear,GIS)局部放电检测分类框架。首先,将局部放电相位(phaseresolvedpartialdischarge,PRPD)分析作为输入序列,利用LSTM学习PRPD信号中的时间关联;其次,将LSTM的结果输入到多头注意力模块中,通过结合多头自注意力和LSTM网络,重点分析与PRPD的不同相位集相对应的不同表征子空间的信息,其中,自注意力机制可识别输入和输出序列之间的重要信息,利用多头自注意力网络捕捉故障PRPD的高阶特征;最后,利用分类层在GIS中进行故障检测。结果结果结果表明:线性SVM的性能最差,传统的机器学习分类算法不能很好地捕捉到数据中的微小差异;CNN+LSTM实现了对多元时序数据的时间依赖性捕捉,性能比SVM方法显著提高;AL+DCNN通过对抗学习框架提高了对数据集不平衡性的处理能力,并改善了提取出特征的通用性;所提方法比AL+DCNN方法的F1分数提升了2.96%,证明结合LSTM和多头注意力机制,能够有效提高GIS中局部放电故障的识别性能,在精度、召回率和F1分数指标上均取得了最好性能。这是因为所提方法通过结合LSTM与注意力网络实现了有效的性能互补,取得了优于其他先进方法的开关柜异常检测性能。结论结论所提方法可以有效检测局部放电故障,有助于确保电力设备的长期稳定运行。

Objectives To achieve long-term stable operation of electrical equipment, a framework for anomaly detection and classification of partial discharge (PD) in gas insulated switchgear (GIS) based on long short-term memory (LSTM) networks and a multi-head attention mechanism was proposed. Methods Firstly, phase resolved partial discharge (PRPD) analysis was used as the input sequence, and LSTM was employed to learn the temporal correlations in PRPD signals. Then, the results from LSTM were fed into a multi-head attention module. By integrating multi-head self-attention with the LSTM network, the frame⁃work focused on different representational subspaces corresponding to different phase sets of PRPD. Sec⁃ondly, the self-attention mechanism identified important information between input and output sequences, while the multi-head self-attention network captured high-order features of faulty PRPD. Finally, a classifi⁃cation layer was utilized for fault detection in GIS. Results The experimental results were as follows: Linear SVM performed the worst, indicating that traditional machine learning classification algorithms were not ef⁃fective in capturing subtle differences in the data. CNN+LSTM achieved temporal dependency capture of multivariate time-series data, significantly improving performance over the SVM method. AL+DCNN en⁃hanced the ability to handle dataset imbalances and improved the generality of extracted features through an adversarial learning framework. The proposed method improved the F score by 2.96% compared to the AL+DCNN method, demonstrating that combining LSTM with a multi-head attention mechanism could effec⁃tively enhance the performance of PD fault identification in GIS. It achieved the best performance in terms of accuracy, recall, and F score. This was because the proposed method achieved effective performance complementarity by combining LSTM with the attention network, outperforming other advanced methods in switchgear anomaly detection. Conclusions The proposed method could effectively detect PD faults, contrib⁃uting to ensuring the long-term stable operation of electrical equipment.