Electric multiple unit (EMU) trains rely on traction rectifiers to convert and regulate power from the catenary supply to onboard traction motors. Conventional control of these rectifiers uses proportional-integral (PI) regulators applied to decoupled direct-quadrature (dq) current components. While this approach is computationally simple, PI tuning requires careful parameter selection, and performance degrades when the train encounters sudden load changes, switching between operational modes, or when component behavior deviates from nominal design assumptions.
This research addresses these limitations by replacing all PI controllers with a single intelligent agent trained using deep reinforcement learning (DRL). The agent learns to compute optimal control decisions across the full operating envelope without explicit mathematical models. The team employed the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, which uses two neural networks to reduce overestimation bias and provides stable learning for continuous control problems.
A critical innovation is the addition of Reward Shaping (RS)—a technique that redesigns the learning signal to guide the agent toward desired behaviors such as DC voltage stabilization and smooth current control. Combined with Prioritized Experience Replay (PER), which focuses learning on historically difficult scenarios, this refinement significantly accelerated convergence and improved multi-condition performance.
Validation was comprehensive: simulation results confirmed stable operation under CRH5 traction duty cycles, while hardware-in-the-loop testing on a real-time platform verified that DRL control matched or exceeded conventional PI performance. Stability was rigorously proven using Lyapunov's second method, providing theoretical assurance of robustness.
The work demonstrates that machine learning can successfully replace classical control structures in safety-critical rail power systems. Future applications may extend to freight trains, urban rail networks, and other high-power rectifier systems where operational diversity and transient performance are critical.
