False Alarm Identification Of Transformer Oil Chromatography Online Monitoring Device Based On Deep Learning

Transformer is an indispensable equipment in power system.Once it is abnormal,it will directly affect the safety of power grid.Online monitoring of oil chromatography is an important tool for real-time monitoring of transformer operating status and can detect abnormal status in a timely manner.However,the inconsistent quality level and imperfect technical standards of oil chromatography online monitoring devices are serious problems,which lead to frequent abnormalities and a large number of false alarms in their practical application in the field.At the same time,it is difficult to verify all alarms by manual methods,and there is no research directly for the identification of false alarms of the device.Therefore,this paper analyzes the online monitoring data of oil chromatography online monitoring devices that have been operating in the field for many years and studies the operation status assessment and false alarm recognition of the device.The paper mainly includes the following content:(1)For the status evaluation of oil chromatography online monitoring device,this paper evaluates the operating status of the device from three perspectives of invalid data distribution,random mutation error and random fluctuation error based on online monitoring data.The invalid data distribution evaluation obtains a quantitative score by counting the proportion of invalid data.Random mutation error evaluation uses the Turkey 53H outlier extraction algorithm to identify data mutation points,and then the quantitative score is based on the number of mutation points and the magnitude of mutation.The EEMD algorithm was used to decompose the online monitoring data into system trend and random fluctuation trend,and artificial noise sequence was constructed based on the system trend.Then the quantitative score was obtained by comparing the signal-to-noise ratio of random fluctuation trend and the artificial noise sequence.Finally,the indexes are synthesized to obtain the device status score and determine whether the device status is abnormal by comparing with the preset threshold.The field application case shows that the method can effectively identify abnormal devices and provide the necessary basic support for improving the accuracy of online monitoring data and the next step of device false alarm identification.(2)For the false alarm recognition of oil chromatographic online monitoring device,this paper uses the trend type of online monitoring data before the alarm to determine whether the alarm is correct.And in order to achieve accurate identification of data trends,a long short-term memory multi-scale attention convolutional neural network(LSTM-MACNN)is proposed.The model first uses the LSTM network to extract long-term dependent features of oil chromatography online monitoring data.Then a residual connection is added,and multi-scale convolution is used to extract complex local features of different scales of the original data and time-dependent features after LSTM processing.Finally,the attention mechanism is used to help the model selectively enhance the important feature information and suppress the irrelevant information.The experimental results show that the accuracy of the LSTM-MACNN model is 96.3%on an H2-total hydrocarbons test set and 98.6%on a C2H2 test set.Example verification shows that the accuracy of identifying false alarms by using the trend type of data before the device alarm is 97%.The proposed method can provide an important reference for the false alarm judgment of oil chromatography online monitoring devices.