Failure rate prediction of low voltage networks using regression models

Abstract: High age of low voltage grids is becoming an increasing issue. This in combinationwith a high penetration of distributed generation will increase the demandfor accurate reliability assessment of the low voltage grid. In comparison withhigh- and medium voltage, the statistical assessment of the LV grid conditionis more or less considered as unspoiled terrain.The aim of this thesis is to investigate and implement statistical model inorder to determine the failure frequency on dierent low voltage components.The components of interest are oil-paper cable, PVC (Polyvinyl chloride) cable,PEX (polyethylene) cable, cable joints and cable terminations. The result ofthis study will enhance the reliability assessment of the low voltage grid thusminimising the outages for customers and outages cost for grid owners.In order to perform this study an extensive data collection procedure isrequired which is carried out mainly with the software program feature manipulationengine (FME). The investigated and utilised statistical models arePoisson regression model, negative binomial regression model, logistic regressionmodel and proportional hazard model (PHM). The results of this study showsthat the factors having the biggest impact on failure frequency are: concessionarea, amount of cable joints, feeding- or service cable, and age. Nevertheless,the other factors: cable type, soil type and total length should not be overlookedsince these can have a signicant impact in local areas.Between the best and worse performing concession area the failure frequencydiers with a factor 3.4. A service cable is 3.3 times more likely to fail comparedto a feeding cable. Between the best and the worst performing age group 0-10years and 70-80 years the dierence is a factor 3.8. One cable joint increase thefailure frequency with a factor 3.1 compared to cables without any cable joints.A oil-paper (FCJJ) cable is 1.6 times more likely to fail compared to a Polyvinylchloride (KKJ) cable.The performance of the dierent statistical models is equivalent for all modelsexcept the Poisson regression model where the performance is worse in allaspects. The remaining models are suitable for dierent data types. For datawith a component that has failed repeatedly is the negative binomial regressionmodel most appropriate. For data where the component of interest is consideredas irreparable can logistic regression be a suitable choice. The proportionalhazard model (PHM) yields a more detailed presentation of how the failure rateis aected over time although it requires failure data with specied time forwhen the failure occurred to be implemented.