计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (10): 300-308.DOI: 10.3778/j.issn.1002-8331.2111-0095

• 工程与应用 • 上一篇    

多尺度特征融合的光伏电站故障诊断

那峙雄,孙涛,来广志,王栋,张长志   

  1. 1.国网电子商务有限公司,北京 100053
    2.国网天津市电力公司 电力科学研究院,天津 300220
  • 出版日期:2022-05-15 发布日期:2022-05-15

Fault Diagnosis for Photovoltaic Power Station by Multi-Scale Features Fusion

NA Zhixiong, SUN Tao, LAI Guangzhi, WANG Dong, ZHANG Changzhi   

  1. 1.State Grid Electronic Commerce Co., Ltd., Beijing 100053, China
    2.Electric Power Science & Research Institute, State Grid Tianjin Electric Power Company, Tianjin 300220, China
  • Online:2022-05-15 Published:2022-05-15

摘要: 为了高效率、低成本地发现光伏电站设备故障,提出了一种多尺度时序特征融合的故障诊断深度学习模型。组合多种尺度的卷积操作,实现了多尺度时序特征提取,能够使网络过滤输入数据的噪点,同时又不丢失数据的关键信息;针对光伏电站监测数据长序列、多维度的特点,采用带有注意力机制的长短时记忆循环网络提取数据中的序列特征;使用了度量学习的损失函数辅助训练,进一步提升模型辨别能力。另外,设计了一种数据规范化算法,能够通过不同设备之间的横向比较和不同日期之间的纵向比较,将监测数据转化为相对度量值,从而矫正设备、天气等因素对监测数据的干扰。实验表明,配合数据规范化算法,该模型在真实的电站监测数据上能够达到95%以上的诊断准确率,明显优于已有的诊断算法,能够满足实际应用场景的故障诊断需求。

关键词: 故障检测, 深度学习, 时序数据分析, 光伏电站

Abstract: To diagnose faulty equipments in time and save the cost of maintenance, a deep learning model of fault diagnosis based on multi-scale time serial features fusion is proposed for the operation and maintenance of photovoltaic power station. To take advantage of the different abilities of convolutional layers with different scales, the model implements the multi-scale neighborhood feature extraction, which can help the network filter the noise in input data while keep the important information. Then, considering the long time series and high dimension of monitoring data, a long short-term memory unit with attention mechanism is used to extract sequential features. Finally, to improve the discernment of the network, metric learning loss function is applied to assist training. In addition, a data normalization algorithm is designed out. It can be used to rectify the impact of devices and weather on monitoring data through comparison between data from different devices and different days. Experiments prove that the proposed model can diagnose faults of equipment more accurately compared with other time series analysis algorithms, which can achieve 95% accuracy on real monitoring data of photovoltaic power station.

Key words: fault diagnosis, deep learning, time series data analysis, photovoltaic power station