[Paper Review] Condition Monitoring of HV Bushings in the Presence of Missing Data Using Evolutionary Computing

Sizwe M. Dhlamini, Fulufhelo V. Nelwamondo+1 more|arXiv (Cornell University)|May 17, 2007

Energy Load and Power Forecasting10 references17 citations

TL;DR

This paper proposes using genetic algorithms (GA) and particle swarm optimization (PSO) to estimate missing dissolved gas analysis (DGA) data in high-voltage (HV) bushings for improved condition monitoring. GA outperformed PSO, achieving 84% accuracy with two missing variables, while PSO dropped to 66%, demonstrating GA's superior robustness in data imputation for bushing condition classification using IEEE C57.104, IEC 599, and IEEE production rate methods.

ABSTRACT

The work proposes the application of neural networks with particle swarm optimisation (PSO) and genetic algorithms (GA) to compensate for missing data in classifying high voltage bushings. The classification is done using DGA data from 60966 bushings based on IEEEc57.104, IEC599 and IEEE production rates methods for oil impregnated paper (OIP) bushings. PSO and GA were compared in terms of accuracy and computational efficiency. Both GA and PSO simulations were able to estimate missing data values to an average 95% accuracy when only one variable was missing. However PSO rapidly deteriorated to 66% accuracy with two variables missing simultaneously, compared to 84% for GA. The data estimated using GA was found to classify the conditions of bushings than the PSO.

Motivation & Objective

• To address the challenge of missing dissolved gas analysis (DGA) data in high-voltage bushing condition monitoring.
• To evaluate the effectiveness of evolutionary computing techniques—specifically genetic algorithms (GA) and particle swarm optimization (PSO)—in estimating missing DGA variables.
• To compare GA and PSO in terms of accuracy and computational efficiency for data imputation in HV bushing classification.
• To improve the reliability of condition classification for oil-impregnated paper (OIP) bushings using imputed DGA data.
• To determine which evolutionary algorithm provides more accurate and stable data estimation under varying missing data conditions.

Proposed method

• The study uses neural networks integrated with particle swarm optimization (PSO) and genetic algorithms (GA) to estimate missing DGA values in HV bushing data.
• PSO and GA are applied to optimize the imputation of missing data points in a dataset of 60,966 oil-impregnated paper (OIP) bushings.
• The classification of bushing conditions is performed using standards-based methods: IEEE C57.104, IEC 599, and IEEE production rates.
• Imputation accuracy is evaluated by comparing estimated values against actual values, with performance measured under one and two missing variables.
• The neural network model is trained using the imputed data to classify bushing conditions, with accuracy assessed across different missing data scenarios.
• Computational efficiency and convergence speed of GA and PSO are compared to evaluate practical deployment feasibility.

Experimental results

Research questions

• RQ1 How accurately can PSO and GA estimate missing DGA data in high-voltage bushings?
• RQ2 How does the performance of PSO and GA degrade when two DGA variables are missing simultaneously?
• RQ3 Which algorithm—PSO or GA—produces more reliable data imputation for HV bushing condition classification?
• RQ4 What is the computational efficiency difference between PSO and GA in the context of missing data estimation?
• RQ5 Does data imputed by GA lead to better classification accuracy of bushing conditions than PSO?

Key findings

• When only one DGA variable was missing, both GA and PSO achieved an average accuracy of 95% in estimating missing values.
• With two variables missing simultaneously, PSO accuracy deteriorated to 66%, while GA maintained 84% accuracy, indicating superior robustness of GA.
• The data estimated using GA resulted in more accurate classification of bushing conditions compared to data estimated using PSO.
• GA demonstrated better stability and reliability in imputing missing data under multiple missing variable scenarios.
• PSO showed faster convergence but poorer performance as missing data complexity increased, highlighting a trade-off between speed and accuracy.
• The study confirms that GA is more suitable than PSO for imputing missing DGA data in HV bushing condition monitoring systems.