Partial Discharge Evolution And Defect Diagnosis For Oil-paper Insulation Under AC-DC Composite Voltage

Converter transformer is one of the key equipment in HVDC transmission system. It provides an efficient solution to short supply of power load. With a rising number of converter transformers being put into operation for prolonged duration, the oil-paper insulation defects have gradually increased,posing a serious threat to the stability of HVDC transmission system. Defect diagnosis for converter transformer is therefore of signifant importance. Compared to AC transformer,converter transformer endures higher temperature, heavierload, as well as more complex internal electric field. Hence, the two types of tranformers demonstrate much difference in terms of the characeteristics on partial discharge and gas dissolved in oil. Given that the existing AC fault diagnosis method is not suitable for converter transformer, it remains imperative to carry out specifically designed research on defect diagnosis for oil-paper insulation under compbined voltage.To simulate the internal operating environment of converter transformer,this paper established a set of PD test platform for oil-paper insulation under composite voltage, which is composed of experimental cavity, heating box,circulatory pump, AC/DC source, as well as the oil smapling point. The platform is capabel of simulating the electric fields of any AC-DC ratio, adjusting oil temperature between 20℃-130℃ as well as precisely controling the oil flow between 0-1.2m3/h. Simulating typical oil paper insulation defects, four test models were designed covering air gap discharge, column plate surface discharge, needle plate surface discharge and turn-to-turn discharge of oil-paper insulation under composite voltage. The platform also realizes the synchronous signal charged detection of pulse current, UHF, ultrasonic and dissolved gas in oil.The tests for four oil paper insulation defect models under different AC and DC proportion, i.e.1:1,1:3,1:5,1:7, were carried out. The phenomena from the inception to the extinction of the discharge were observed. The coorelation between discharge and the proportion of composite voltage was analysed. It was found that, with the increase of DC voltage components, partial discharge inception voltage rises gradually, the number of partial discharge reduces, partial discharge magnitude decreases, dissolved gas in oil composition decreases, partial discharge phase distribution area reduces, the discharge point gradually scatters, phase width narrows down. The regular pattern of oil-paper discharge under composite voltage is sumarized as that at the innitial phase,discharge is faint;at the develop phase, it has a slight enhancement; at the dangerous phase, the strength of discharge hits the top.To realize the oil-paper fault diagnosis under composite voltage,48 discharge types were clasiified and studied according to different AC and DC proportion, defect types and discharge severity levels. The paper divides the discharge spectra into 4 types:statistical diagram, the waveform scheme, chromatographic images, as well a wholistic spectra including all above.72,13,17 and 102 fingerprint features were extracted from the four types of spectra respectively. The correct rate for pattern recognition using the original fingerprint features is only 37.8%.To improve the precision of identification, two optimization methods were studied for dimensionality reduction. The K-L transformation is based on signal classification, obeying the rule of the new feature space minimum error after dimensionality reduction. With the support of K-L transformation, the number of the fingerprint features in the four types of spectra were reduced to 9,5,8 and 4 respectively. As a result, the optimized test improved the correct rate by 26.6%. The Rough Set Theory is based on classification mechanism,analying the support of decision attribution for type to reduce dimension. Using rough set theory, the number of the fingerprint features in the four types of spectrawere reduced to 7,5,12 and 24 respectively improveing the correct rate by 27.1%. Both methods achieve good effect of optimization, and the optimization effect is close.To get the efficient method for oil-paper discharge diagnosis, direct recognition for 48 discharge patterns were firstly carried out to refelct different AC and DC proportions, four defect types and three severity levels at one time. According to the comparision of five discharge pattern recognition methods, repeat clips, fuzzy recognition, random forests(RF), support vector machine and BP neural network, an appropriate way of derect recognition for 48 kinds of discharge is given, which is to choose the wholistic spectra with RST optumization. Using this means, the correct rate of the test mounts up to 85.4%.To improve the correct rate of recoginition, fractional step pattern recognition is put forward,which identifies AC/DC proportions,defect types and severity levels step by step. Moever, with the method of RF, the correct rate of the test reaches as high as 90.4% during a rather prolonged time span of 630s.In addition to improve correct rate within a shortened time period at the same time, this paper adopts diverse indicators to realize identification under different circumstances. Chromatographic spectra,with 8 fingerprint features in it after K-L transformation were firstly excerted random forest test to recognize AC/DC proportion and defect types. Repeat clips method was then used foridentification of severity to ensure the correct rate and speed of the test. It is found that the comnied methods of RF-RF-MULITIEDT is more efficent with its correct rate of test reaching 89.7% and the he recognition time decreaing from 630s to 150s at the same time.In conclusion, this paper identifies an appropriate way to recognise the pattern of oil-paper discharge. Using chromatographic library with the K-L transformation, RF and repeat clips methods to recognize step by step the proportion of voltage,type of defect, level of severity, fast and acurate diagnosis is realized with the correct rate being 89.7% and recognition time being 150s, which provide referntial criteria for the diagnosis of converter tranformer defects.