An Improved Deep Reinforcement Learning Control Strategy for Traction Dual Rectifiers in EMUs

Jul 16, 2026

Addresses control-performance sensitivity in CRH5 traction dual rectifiers caused by PI-based dq current decoupling, where linearized methods may degrade under reference changes or model mismatch and nonlinear methods can introduce jitter and weak steady-state accuracy. Proposes replacing all PI elements with a single Deep Reinforcement Learning (DRL) intelligent agent to stabilize intermediate DC-link voltage. Identifies TD3 limitations across varying operating and switching conditions, and improves rewards via Reward Shaping combined with Prioritized Experience Replay for faster convergence. Validates effectiveness through simulation, Lyapunov second-method stability analysis, and hardware-in-the-loop (HIL) results.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages19

Reading Time48 min

Introduction

Issue Description

Literature Review

An Improved Deep Reinforcement Learning Control Strategy for Traction Dual Rectifiers in EMUs Zhigang Liu, Mingwei Tang, Xiangyu Meng, Hui Wang, Qiao Zhang, Haoyu Wang, Mengru Li arXiv :2607 .09276v1 [ ee ss . SY] 10 Jul 2026 Abstract—Due to the use of PI-based dq current decoupling in the pulse rectifier of CRH5 high-speed trains, the PI parameters directly affect the traction system’s control performance. Linearized control may have issues with reference trajectory changes or model mismatches, leading to a decrease in system performance, while nonlinear control may have problems with jitter and poor steady-state accuracy. This paper proposes anew control strategy that replaces all PI in the dq current decoupling control with a single intelligent agent. This method based on Deep Reinforcement Learning (DRL) can avoid various drawbacks of linearization and nonlinear control and ensure the stability of intermediate DC voltage. However, when EMUs are in different working conditions and switching, the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm used in traction dual rectifiers does not have a good control effect. Focusing on the issue, Reward Shaping (RS) is added to redesign a nonlinear reward function, which can be combined with Prioritized Experience Replay (PER) to increase the convergence speed of the episode reward. The simulation results show that the improved control strategy can be effectively applied to EMUs working in multiple conditions. Finally, the stability analysis is carried out using Lyapunov’s second method and the verification results of the hardware-in-the-loop (HIL) simulation platform show that the DRL control has a good effect. Index Terms-EMUs, dq decoupling control, Deep Reinforcement Learning, traction dual rectifiers, multi-working conditions, hardware-in-the-loop (HIL) . I. INTRODUCTION At present, the power unit of China’s EMUs mainly uses AC-DC-AC traction transmission system, which is mainly composed of pulse rectifier, DC circuit, traction inverter, traction motor and so on. All links interact with each other, especially the converters and other power electronic devices have strong nonlinearity and impact, which eventually cause the power quality problem of the coupled system of EMUsand traction network in the different operating conditions of EMUs [1] . Among them, the pulse rectifier needs to maintain the stability of the DC link voltage to provide good conditions for the inverter through the control of the rectifier. A. ISSUE DESCRIPTION The current pulse rectifier combined with the control strategy adopts sine pulse width modulation (SPWM) technology to control the on/off of IGBT, achieving real-time control of various electrical quantities in the rectifier circuit. From the development of traditional PID control, modern controls such as variable structure control, robust control, predictive control, etc., and intelligent controls such as fuzzy control, expert control, and neural network control have emerged, which do not rely on the mathematical model of the system. In this paper, the dq decoupling control in CRH5 EMUs is taken as an example to analyze. The double closedloop control structure based on PI has the bandwidth problem of voltage loop and current loop[2], which will affect the stability of the rectifier circuit. At the same time, the selection of multiple PI parameters in the voltage and current loops is related to the dynamic and static performance of the rectifier output. Inappropriate values may exacerbate the negative impedance characteristics of the rectifier and even lead to lowfrequency oscillations [2] . Due to the nonlinear characteristics of converters, the feedback linearizing control [4] and model predictive control [5] need to rely on accurate mathematical models when linearizing the system, which may lead to the deterioration of system performance due to the change of reference trajectory or model mismatch [6] . Since these nonl

FAQs

Why do PI parameters in dq current decoupling affect the traction system’s control performance?

The paper states that PI parameters directly influence control performance in CRH5 pulse rectifier traction systems, and incorrect parameter selection can worsen negative impedance characteristics, potentially causing low-frequency oscillations and stability issues.

What control strategy does the paper propose instead of using PI controllers?

It replaces all PI components in the dq current decoupling control with a single DRL-based intelligent agent to avoid drawbacks of purely linearized or nonlinear control and to ensure stability of the intermediate DC-link voltage.

How are TD3 limitations handled under different EMU working conditions and switching?

The paper adds Reward Shaping to redesign the nonlinear reward function, and combines it with Prioritized Experience Replay to improve convergence speed of episode reward and achieve better control across multiple conditions.

Jul 16, 2026

Upload by Evangeline