Reliability and Ageing Assessment of XLPE Power Cables Using Bayesian Neural Network

With the global growth of electrification, there is an increased reliance on underground power cable networks. Due to new electrical loads, e.g., Data Centers, Heat Pumps, etc., this significant increase and variation in distribution and patterns of demand, utilities face escalating challenges in ensuring the long-term reliability of ageing crosslinked polyethylene (XLPE) power cables,. Traditional condition assessment approaches rely heavily on deterministic indicators that fail to account for uncertainty in degradation processes, often leading to conservative or suboptimal maintenance decisions. Moreover, existing diagnostic-based models typically provide single-point health estimates, offering limited insight into confidence levels and risk, which constrains effective reliabilitycentered asset management. This paper proposes a probabilistic framework for enhancing the reliability of XLPE power cables using a Bayesian Neural Network (BNN). Unlike traditional deterministic models that provide point estimates of asset health, the proposed BNN explicitly quantifies the uncertainty inherent in cable degradation. The framework utilizes a dataset of 5732 real power cable samples, incorporating 7 critical diagnostic indicators: operational age, partial discharge, tangent delta characteristics, visual condition, and neutral corrosion. The model targets the health index (HI) as a continuous output variable. Model architecture is optimized through a comprehensive sensitivity analysis for hyperparameter tuning, evaluating the impact of prior distributions and network depth on predictive stability. Generalization capability is evaluated by excluding a 5 % subset of the dataset, which served as unseen samples for final validation. The BNN outputs a posterior predictive distribution for each power cable, yielding a mean predicted HI and a 95 % confidence interval. The model achieves R²=0.991. The impact of this study lies in transforming power cable maintenance and asset management planning from a reactive to a risk-based paradigm. By providing uncertainty bounds, the model enables utility operators to distinguish between confident safe and uncertain/high-risk assets. This probabilistic output facilitates precise maintenance scheduling and optimized lifecycle management, ensuring interventions are targeted based on both the predicted condition and the confidence of that prediction.

Keywords: Asset management, Bayesian neural network, Health condition, Lifecycle management, Maintenance scheduling, Reliability.