Transformer health index assessment based on neural- fuzzy method utilising monte carlo simulation

Transformers can be subjected to multiple types of stresses which could reduce their reliability under long service period. Since transformers are one of the important equipment in power systems, it is important to monitor its condition in order to avoid unnecessary failures and this can be done through a condition based management. Normally, the condition of transformers is evaluated through a single quantitative indicator known as Health Index (HI). Conventionally, HI is determined by scoring method that based on historical information of transformers population and expert judgement. Alternatively, Artificial Intelligence (AI) techniques like Fuzzy Logic (FL) and Artificial Neural Network (ANN) were proposed to overcome these drawbacks. However, these techniques suffer from complexity of producing the inference rules of FL and difficulty of choosing the appropriate ratio of training data for ANN. In this research, the aim is to apply an alternative method to determine the HI of transformers based on Neural-Fuzzy network (NF) method that can overcome the issues in previous AI and scoring methods. Two schemes were implemented to train the NF network which were based on in-service condition data and Monte Carlo Simulation (MCS) data. The conventional scoring method was also applied for comparison purpose. The performances of these methods were tested on two case studies which had transformers with voltage level less than 69 kV. In-service condition data such as furans, dissolved gases, moisture, AC Breakdown Voltage (ACBDV), dissipation factor (DF), acidity, interfacial tension (IFT), colour and age were fed as input parameters to the NF network. Multiple studies were carried out to test the performance of NF on HI of transformers which included the effects of training data number, age, dissolved gases and in-service condition data. It is found that the ratio of 80% training and 20% testing is sufficient for NF trained by in-service condition data method. For NF trained by MCS data method, the optimum number of training data required is 1000. The introduction of age in the NF method provides additional input for assessment of transformers. The NF trained by MCS data has no issue adapting with Total Dissolved Combustible Gases (TDCG) as input data. However, NF method requires a minimum number of in-service condition input data in order to carry out practical assessment on transformers condition. In general, compared to the other two methods, NF trained by MCS data method can provide a realistic alternative assessment of transformers. This technique can be used to diagnose the condition of transformers without the reliance on the historical information of transformers population and expert judgment.

Preview Text FK 2016 91 - IR.pdf Download (1MB) | Preview

Actions (login required)

View Item