Need for AI in Transformer Diagnostics and Prognostics

In this paper we present the Probabilistic Operations Warranted for Energy Reliability Evaluation and Diagnostics (POWERED) hybrid artificial intelligence (HAI)/ machine learning (ML) tool for diagnosing and predicting performance and remaining useful life (RUL) of electrical transformers to increase reliability and inform maintenance. POWERED incorporates electrical, thermal, and environmental factors that influence transformer degradation to perform predictive analytics and forecast their health and status (H&S) indicators (internal temperatures, dissolved gases) to predict impending failures and RUL more accurately. POWERED uses a rich and modular probabilistic modeling approach that holistically integrates various types of models including sub-symbolic (i.e., purely data-driven) models and symbolic (i.e., Bayesian and physics-based) models. POWERED's predictive analytics are demonstrated on a rich data set collected over one year from an operational distribution transformer. The data set includes timestamped ambient temperatures, oil and hot spot winding temperatures, and electrical loading and concentrations of dissolved gases. We show how these diverse factors can be correlated and used both in real-time H&S monitoring and what-if analysis under extreme conditions.