Dynamic control optimization and steady-state performance modeling of switching power supply based on deep reinforcement learning

Switching power supplies exhibit slow response, significant overshoot, and limited steady-state efficiency under dynamic operating conditions. This paper proposes a method for dynamic control optimization and steady-state performance modeling based on deep reinforcement learning (DRL). Taking a phase-shifted full-bridge (PSFB) soft-switching circuit as the subject, a high-fidelity simulation platform integrating small-signal modeling and DRL training environments is constructed. A TD3 control framework with a 12-dimensional state space and continuous duty cycle action space is designed, incorporating an innovative hierarchical reward mechanism: The dynamic phase optimizes response time and overshoot, while the steady-state phase constrains efficiency and total harmonic distortion (THD), achieving multi-objective co-optimization of dynamic response and steady-state performance. Simulation results demonstrate that under typical operating conditions—including load transients, input fluctuations, and reference steps—the proposed DRL controller achieves overshoot below 0.8%, reduces settling time to 0.9 ms, maintains steady-state error within ±5 mV, and outperforms traditional PID and Linear Adaptive Disturbance Rejection Control (LADRC) methods in efficiency. This validates its effectiveness and superiority for intelligent control of power electronic devices.