Efficient decision-making and controlling of the primary and secondary assets involved in the grid enables the accomplishment of reduced downtime and condition- based favorable outcome by bringing the whole grid one step closer to the autonomous future one. This concept can be achieved by applying Predictive Maintenance (PdM). PdM is the asset maintenance strategy that uses Machine Learning (ML) to adjust operating conditions for desired outcomes, as well as intelligently schedule asset maintenance. In other words, it adds the ability to give advice to the technician on what and how to do the repair by taking advantage of artificial intelligence (AI). Admittedly, this issue is mostly related to the pillar of the grid - transformers. Prior to the maintenance, it is essential to grasp the health state of the power transformer parts (components) and then plan accordingly. To determine the overall transformer behavior in mid-term and long-term perspectives, several tests and factors are taken into consideration. Those considerations are quantified and transformed into the numerical and categorical values, so-called Health Index (HI). This thesis specifically focuses on research and developing an HI estimation model for the power transformers which would determine the current HI and forecast the future HI throughout the power transformer lifespan for the customers to make decisions on scheduling the maintenance and planning the risk costs by leveraging the historical information on the gas sensor values, winding and oil temperature, cooler state, loading, and other classical transformer parameters. It is achieved by applying hybrid approach: (1) industrial standards and experts’ judgements and (2) AI-driven condition models. To validate it, we first discuss different conventional HI evaluation methods as well as novel AI models. Afterwards, analytical approach of the HI estimation based on weighted-sum method is conducted and customized HI estimation framework is provided. Following this, new employed ML techniques are compared based on the metrics of errors and the best one is deployed. To verify the proposed approach, both the aging curve model and ML algorithms are tested on two 500kV transformers dataset with 32 parameters. The final AI-driven HI estimation approach establishes a strong basis for deploying ML with the customized condition model and a reliable real-time decision- support for the substation operators.
