基于时间卷积网络的台区总表故障智能辨识技术

Fault Identification of Main Meters Based on Temporal Convolutional Nets

台区总表关联10 kV线路售电侧电量及台区供电侧电量,起到承上启下的作用。台区总表故障导致的电能量异常,会同时对10 kV及台区线损的计算产生影响,造成线损合理率下降,不利于供电可靠性。对台区总表故障进行有效辨识是解决以上问题的关键。对此提出基于时间卷积网络的台区总表故障辨识技术,通过提取台区量测值时间序列中的高维特征实现对故障的辨识。进一步地,在年度序列故障辨识中引入注意力机制,以甄别出关键的信息片段,提升辨识准确率。上海市10 kV变压器台区的算例结果表明,所提出的故障辨识算法可实现93.1%的日度故障辨识精度以及85.3%的年度故障辨识精度,且相比传统深度学习模型具有更高的计算效率。

Main meters in 10 kV level connect the supply side and the retail side and record the electricity delivery.The abnormal electrical energy caused by the failure of the main meter in the substation area will have an impact on the calculation of 10 kV and substation area line losses,resulting in a decrease in the reasonable rate of line losses,which is not conducive to power supply reliability.Therefore,fault identification of transformer meters is the key to avoid such risk.In this paper,a fault identification method for main meters based on Temporal Convolutional Nets,which can identify the fault by extracting high-level features from the time series of the transformer meter measurements,is proposed.Attention mechanism is further introduced to identify the key data segments in yearly fault identification.Case study based on 10 kV transformers in Shanghai shows that our method can achieve 93.1%accuracy in daily fault identification and 85.3%in yearly fault identification,and has better computation efficiency than traditional deep-learning models.