| Transformer oil is an important insulating medium in power transformers.Water content in oil is an important factor to determine the performance of the transformer.Many problems in the insulation system,such as the decrease of breakdown voltage,the increase of dielectric loss,the acceleration of organic chemical reaction,and even the major accidents such as insulation breakdown and burnout of equipment,are caused by the higher water content in oil.However,at present,most of the electric power industry takes the insulating oil out of the running transformer on schedule and measures its water content in the external laboratory by Karl Fischer titration method.Although this method has high detection accuracy,it can not find the potential fault of the transformer in time.Therefore,researching a timely and effective method for detecting the water content in transformer oil is of great significance for ensuring the safe and stable operation of the transformer.Ultrasonic testing technology is a kind of technology based on the attenuation characteristics of the acoustic propagation medium to obtain the characteristic parameters of the propagation medium.In this paper,the multi frequency ultrasonic testing equipment of transformer oil is used to detect the attenuation characteristics of ultrasonic in oil by the combination of penetration testing and reflection testing,and the 242 dimensional multi frequency ultrasonic data which can characterize the characteristic information of water content in transformer oil were obtained,and the relationship between the multi frequency ultrasonic data and the change of micro water content in transformer oil was studied and analyzed.PCA-GA-BPNN was used to establish the regression identification model of the water content in transformer oil,and finally the effective detection of the water content in the oil was realized.The main research work and results are as follows:(1)The system adopted in this paper combines ultrasonic penetration detection method and ultrasonic reflection detection method to conduct multi-frequency ultrasonic detection of transformer oil,and its detection center frequency is 750 kHz and the detection frequency range is 600kHz-1000 kHz.The signal of the multi frequency ultrasonic L1 phase is a reference signal,and the amplitude response of theL1 phase is positively correlated with the water content in oil.The amplitude response of the L2 phase is inversely related to the water content in oil,and the higher the water content in oil,the greater the difference in the amplitude response of the L1 and L2 phases of the same oil sample at the same frequency point.There is no significant positive and negative correlation between the amplitude response of L3 phase multi frequency ultrasound and the water content in the oil at all detection frequencies.Except for a few special cases,when the water content in the oil exceeds 25mg/L,the amplitude of the L3 signal at 943.7kHz is less than 0.05 V.(2)Two dimensionality reduction algorithms PCA and MDS are used to reduce the original 242 dimensional ultrasonic data.The accuracy of the two dimensionality reduction algorithms is verified under the two models of back propagation neural network and generalized regression neural network.The results show that both PCA and MDS dimensionality reduction algorithms can achieve effective dimensionality reduction of the original multi frequency ultrasound data,and when the data dimension is 23 dimensions,the accuracy of PCA is 99% and 97% better than MDS in the BPNN and GRNN,respectively.(3)Regression identification models of transformer oil micro-water content based on PCA-GA-BPNN and PCA-PSO-GRNN were established.Under the same training and test set conditions,the PCA-GA-BPNN and PCA-PSO-GRNN water content in transformer oil regression identification models are compared and analyzed.The results show that the accuracy of PCA-GA-BPNN model is 92.93%,and the accuracy of PCA-PSO-GRNN model is 89.69%.In addition,the MAPE(7.07%),RMSE(0.2339)and perr(0.0039)of PCA-GA-BPNN are better than PCA-PSO-GRNN.It can be known that PCA-GA-BPNN model is better than PCA-PSO-GRNN model in predicting the water content in transformer oil. |
