Reinforcement learning-based controlled switching approach for inrush current minimization in power transformers

Abstract

Transformers are essential components for the reliable operation of the power grid and the connection of renewable energy sources (RES) to them. The transformer core is constituted by a ferromagnetic material and depending on the magnetization state of the core, the energization of the transformer can lead to high magnetizing inrush currents. Such high amplitudes can shorten the transformer life expectancy and cause power quality issues in the power grid, thereby challenging the reliable integration of RES to the grid. Various Machine Learning (ML) methods have been proposed to classify inrush currents. However, minimizing inrush current implies learning a sequence of actions considering the dynamic operation of the transformer. This is different from a classification problem and entails substantial challenges associated with magnetization properties of materials and power transformer operation dynamics. This paper introduces a novel Reinforcement Learning (RL) approach for inrush current minimization, capable of generating sequential decision-making strategies adapted to the dynamic operation of power transformers. The proposed method employs the Proximal Policy Optimization (PPO) algorithm and is trained and evaluated using an equivalent duality-based model of a real 7.4 megavolt–ampere (MVA) power transformer. The PPO-based strategy is compared with two benchmarks: Deep Q-Network (DQN) and the classical circuit breaker (CB) switching method used in laboratory tests. Results demonstrate that the PPO approach can minimize inrush current and achieves a reduction of 14.29% compared to DQN and 79.78% compared to laboratory measurements.