An Imputation Reinforcement Learning Agent For Power Transformers Load Study

Among the challenges occurring when performing power transformers load study is the lack of data stemming from the loss of the heat-run factory test report, or the inability to properly collect the required inputs. To address the issue, this paper introduces a novel imputation strategy for the required attributes values. By adopting a data-driven approach, a reinforcement learning agent with little prior knowledge and using a multivariate random walks simulation of the state space is devised to navigate a thermal environment where unknown temperature rises are estimated against the normal life expectancy loading scenario. A performance evaluation of the imputation agent with a case study evidencing a natural cooled air transformer showcases prediction error within 5% range, signaling the efficacy of the imputation agent.