Reinforcement Learning-Based Controlled Switching Approach for Inrush Current Minimization in Power Transformers

TL;DR

This work tackles transformer energization inrush by marrying a duality‑based transformer model (incorporating a Jiles–Atherton core) with reinforcement learning to learn the optimal circuit breaker closing instant from states consisting of the opening angle and remanent fluxes $\phi_1,\phi_2,\phi_3$. The RL agents (DQN variants and PPO) are trained in a simulated environment and evaluated on real data, with a reward that penalizes peak inrush $I_{max}$ when it exceeds 1 pu. Results show PPO achieving the most stable and effective inrush minimization, reducing average peak currents by roughly 76–80% relative to non‑controlled closures and keeping peaks under 1 pu in most cases. The study demonstrates the viability of RL‑driven controlled switching for inrush mitigation and highlights practical considerations for deployment, including the need for fast surrogate models for system‑level studies.

Abstract

Transformers are essential components for the reliable operation of power grids. The transformer core is constituted by a ferromagnetic material, and accordingly, depending on the magnetization state, the energization of the transformer can lead to high magnetizing inrush currents. Such high amplitudes shorten the life expectancy of a transformer and cause power quality issues in power grids. Various techniques have been proposed to minimize the inrush current; however, the application of Reinforcement Learning (RL) for this challenge has not been investigated. RL incorporates the ability to learn inrush minimization strategies adjusted to the dynamic transformer operation environment. This study proposes an inrush current minimization framework by combining controlled switching with RL. Depending on the opening angle of the circuit breaker and the remanent fluxes at disconnection, the proposed method learns the optimal closing instant of the circuit breaker. Two RL algorithms have been trained and tested through an equivalent duality-based model of a real 7.4 MVA power transformer. The evaluation of the RL algorithms is carried out with real measurement data and compared with real laboratory inrush currents. The results show that the inrush current is effectively minimized with the proposed framework.

Figures (12)

• Figure 1: Duality-based transformer electrical circuit of a three-phase transformer Chiesa2010_2.
• Figure 2: Agent-environment interaction in RL Sutton2015.
• Figure 3: Overall block diagram of the proposed approach.
• Figure 4: Diagram of the environment.
• Figure 5: Real remanent flux data, fitted curve and 95$\%$ CI.